refactor(benchmark): 整理避障压测逻辑和报告

FishROV/Assets/Scripts/Benchmark/Agents/BenchmarkAgent.cs

@@ -9,8 +9,10 @@ namespace FishROV.AvoidanceBenchmark
         private static Mesh sharedConeMesh;
         private static Material sharedConeMaterial;
         private static MaterialPropertyBlock colorPropertyBlock;
+        private static bool useCustomConeMaterial;
 
         private Vector3 velocity;
+        private Vector3 previousVelocity;
         private float maxSpeed = 1f;
 
         public BenchmarkAgentRole Role { get; private set; }
@@ -18,8 +20,18 @@ namespace FishROV.AvoidanceBenchmark
         public float Radius { get; private set; }
         public float MaxSpeed => maxSpeed;
         public Vector3 Velocity => velocity;
+        public float CurrentSpeed { get; private set; }
+        public float PeakSpeed { get; private set; }
+        public float CurrentAcceleration { get; private set; }
+        public float PeakAcceleration { get; private set; }
         public Vector3 Target { get; set; }
 
+        public static void SetSharedConeMaterial(Material material)
+        {
+            useCustomConeMaterial = material != null;
+            sharedConeMaterial = material;
+        }
+
         public static BenchmarkAgent CreatePrimitive(BenchmarkAgentSpawnInfo spawnInfo, Transform parent)
         {
             var go = CreateConeObject();
@@ -50,6 +62,12 @@ namespace FishROV.AvoidanceBenchmark
             GroupId = spawnInfo.GroupId;
             Radius = spawnInfo.Radius;
             maxSpeed = Mathf.Max(0.01f, spawnInfo.MaxSpeed);
+            velocity = Vector3.zero;
+            previousVelocity = Vector3.zero;
+            CurrentSpeed = 0f;
+            PeakSpeed = 0f;
+            CurrentAcceleration = 0f;
+            PeakAcceleration = 0f;
             ApplyColor();
         }
 
@@ -58,8 +76,20 @@ namespace FishROV.AvoidanceBenchmark
             velocity = Vector3.ClampMagnitude(preferredVelocity, maxSpeed);
         }
 
+        public void SetMaxSpeed(float value)
+        {
+            maxSpeed = Mathf.Max(0.01f, value);
+            velocity = Vector3.ClampMagnitude(velocity, maxSpeed);
+        }
+
         public void Integrate(float deltaTime, AvoidanceDimensionMode dimensionMode, float arenaRadius)
         {
+            CurrentSpeed = velocity.magnitude;
+            PeakSpeed = Mathf.Max(PeakSpeed, CurrentSpeed);
+            CurrentAcceleration = deltaTime > 0f ? (velocity - previousVelocity).magnitude / deltaTime : 0f;
+            PeakAcceleration = Mathf.Max(PeakAcceleration, CurrentAcceleration);
+            previousVelocity = velocity;
+
             var next = transform.position + velocity * deltaTime;
             if (dimensionMode != AvoidanceDimensionMode.XYZ3D)
             {
@@ -91,6 +121,11 @@ namespace FishROV.AvoidanceBenchmark
                 return;
             }
 
+            if (useCustomConeMaterial)
+            {
+                return;
+            }
+
             var color = Color.white;
             switch (Role)
             {
@@ -107,6 +142,7 @@ namespace FishROV.AvoidanceBenchmark
 
             var propertyBlock = GetColorPropertyBlock();
             propertyBlock.SetColor("_Color", color);
+            propertyBlock.SetColor("_BaseColor", color);
             renderer.SetPropertyBlock(propertyBlock);
         }
 
@@ -173,13 +209,64 @@ namespace FishROV.AvoidanceBenchmark
                 return sharedConeMaterial;
             }
 
-            sharedConeMaterial = new Material(Shader.Find("Standard"))
+            useCustomConeMaterial = false;
+            sharedConeMaterial = FindFirstBuiltinMaterial(
+                "Default-Material.mat",
+                "Default-Diffuse.mat");
+            if (sharedConeMaterial != null)
+            {
+                return sharedConeMaterial;
+            }
+
+            var shader = FindFirstAvailableShader(
+                "Standard",
+                "Universal Render Pipeline/Lit",
+                "Universal Render Pipeline/Simple Lit",
+                "Universal Render Pipeline/Unlit",
+                "Sprites/Default",
+                "UI/Default");
+
+            if (shader == null)
+            {
+                Debug.LogError("BenchmarkAgent failed to find a built-in shader for runtime cone agents.");
+                return null;
+            }
+
+            sharedConeMaterial = new Material(shader)
             {
                 name = "Benchmark Agent Shared Material"
             };
             return sharedConeMaterial;
         }
 
+        private static Material FindFirstBuiltinMaterial(params string[] materialNames)
+        {
+            for (var i = 0; i < materialNames.Length; i++)
+            {
+                var material = Resources.GetBuiltinResource<Material>(materialNames[i]);
+                if (material != null)
+                {
+                    return material;
+                }
+            }
+
+            return null;
+        }
+
+        private static Shader FindFirstAvailableShader(params string[] shaderNames)
+        {
+            for (var i = 0; i < shaderNames.Length; i++)
+            {
+                var shader = Shader.Find(shaderNames[i]);
+                if (shader != null)
+                {
+                    return shader;
+                }
+            }
+
+            return null;
+        }
+
         private static MaterialPropertyBlock GetColorPropertyBlock()
         {
             return colorPropertyBlock ??= new MaterialPropertyBlock();

FishROV/Assets/Scripts/Benchmark/AvoidanceBenchmarkLogWriter.cs

@@ -37,6 +37,7 @@ namespace FishROV.AvoidanceBenchmark
         public void Record(
             AvoidanceMetricsRecorder metrics,
             AvoidanceCollisionProbe collisionProbe,
+            AvoidancePerfBootstrap.MotionStats motionStats,
             int agentCount)
         {
             if (!hasRun || Time.unscaledTime < nextSampleTime || samples.Count >= MaxSamples)
@@ -55,7 +56,11 @@ namespace FishROV.AvoidanceBenchmark
                 GcAllocKb = metrics.GcAllocBytes / 1024.0,
                 HorizontalOverlapPairs = collisionProbe.HorizontalOverlapPairs,
                 SpatialOverlapPairs = collisionProbe.SpatialOverlapPairs,
-                MinimumClearance = collisionProbe.MinimumClearance
+                MinimumClearance = collisionProbe.MinimumClearance,
+                AverageSpeed = motionStats.AverageSpeed,
+                MaxSpeed = motionStats.MaxSpeed,
+                AverageAcceleration = motionStats.AverageAcceleration,
+                MaxAcceleration = motionStats.MaxAcceleration
             });
         }
 
@@ -72,7 +77,7 @@ namespace FishROV.AvoidanceBenchmark
 
             using (var writer = new StreamWriter(path, false))
             {
-                writer.WriteLine("framework,scenario,dimension,fishCount,arenaRadius,screenWidth,screenHeight,targetFrameRate,timeSeconds,agentCount,averageFps,onePercentLowFps,simulationMs,gcAllocKb,horizontalOverlapPairs,spatialOverlapPairs,minimumClearance");
+                writer.WriteLine("framework,scenario,dimension,fishCount,arenaRadius,screenWidth,screenHeight,targetFrameRate,timeSeconds,agentCount,averageFps,onePercentLowFps,simulationMs,gcAllocKb,horizontalOverlapPairs,spatialOverlapPairs,minimumClearance,averageSpeed,maxSpeed,averageAcceleration,maxAcceleration");
                 for (var i = 0; i < samples.Count; i++)
                 {
                     var sample = samples[i];
@@ -109,6 +114,14 @@ namespace FishROV.AvoidanceBenchmark
                     writer.Write(sample.SpatialOverlapPairs);
                     writer.Write(',');
                     WriteFloat(writer, sample.MinimumClearance);
+                    writer.Write(',');
+                    WriteFloat(writer, sample.AverageSpeed);
+                    writer.Write(',');
+                    WriteFloat(writer, sample.MaxSpeed);
+                    writer.Write(',');
+                    WriteFloat(writer, sample.AverageAcceleration);
+                    writer.Write(',');
+                    WriteFloat(writer, sample.MaxAcceleration);
                     writer.WriteLine();
                 }
             }
@@ -151,6 +164,10 @@ namespace FishROV.AvoidanceBenchmark
             public int HorizontalOverlapPairs;
             public int SpatialOverlapPairs;
             public float MinimumClearance;
+            public float AverageSpeed;
+            public float MaxSpeed;
+            public float AverageAcceleration;
+            public float MaxAcceleration;
         }
     }
 }

FishROV/Assets/Scripts/Benchmark/AvoidancePerfBootstrap.cs

@@ -12,6 +12,13 @@ namespace FishROV.AvoidanceBenchmark
         [SerializeField] private AvoidanceDimensionMode dimensionMode = AvoidanceDimensionMode.XZ25D;
         [SerializeField] private BenchmarkViewMode viewMode = BenchmarkViewMode.Overview;
         [SerializeField] private int fishCount = 300;
+        [SerializeField] private Material fishMaterial;
+        [SerializeField] private float fishMaxSpeed = 2.4f;
+        [SerializeField] private float baitDemoMaxSpeed = 4.2f;
+        [SerializeField] private float speedMultiplier = 1f;
+        [SerializeField] private float timeScale = 1f;
+        [SerializeField] private float hookDropHeight = 10f;
+        [SerializeField] private float hookDropDuration = 1.5f;
 
         private readonly List<BenchmarkAgent> agents = new List<BenchmarkAgent>(512);
         private readonly AvoidanceScenarioController scenarioController = new AvoidanceScenarioController();
@@ -21,17 +28,25 @@ namespace FishROV.AvoidanceBenchmark
         private AvoidanceBenchmarkConfig config;
         private IAvoidanceAdapter adapter;
         private Transform agentRoot;
+        private Transform baitMarker;
         private Camera benchmarkCamera;
         private bool paused;
         private int pendingFishCount;
+        private MotionStats motionStats;
+        private bool hookDropping;
+        private float hookDropElapsed;
+        private Vector3 hookDropStart;
+        private Vector3 hookDropTarget;
 
         private void Start()
         {
             QualitySettings.vSyncCount = 0;
             Application.targetFrameRate = -1;
+            ApplyTimeScale();
             pendingFishCount = fishCount;
             metrics.Start();
             EnsureCameraAndLight();
+            EnsureBaitMarker();
             ResetBenchmark();
         }
 
@@ -46,6 +61,9 @@ namespace FishROV.AvoidanceBenchmark
             }
 
             metrics.BeginSimulation();
+            UpdateHookDrop(Time.deltaTime);
+            ApplyRuntimeSpeed();
+            scenarioController.Tick(Time.deltaTime);
             foreach (var agent in agents)
             {
                 var velocity = scenarioController.GetPreferredVelocity(agent, Time.deltaTime);
@@ -58,24 +76,40 @@ namespace FishROV.AvoidanceBenchmark
                 agent.Integrate(Time.deltaTime, config.DimensionMode, config.ArenaRadius);
             }
 
+            motionStats = CalculateMotionStats();
             collisionProbe.Update(agents, config.DimensionMode, Time.unscaledDeltaTime);
             metrics.EndSimulation();
-            logWriter.Record(metrics, collisionProbe, agents.Count);
+            logWriter.Record(metrics, collisionProbe, motionStats, agents.Count);
         }
 
         private void OnGUI()
         {
-            GUILayout.BeginArea(new Rect(12, 12, 390, Screen.height - 24), GUI.skin.box);
+            GUILayout.BeginArea(new Rect(12, 12, 430, Screen.height - 24), GUI.skin.box);
             GUILayout.Label("FishROV 本地避障压测");
             GUILayout.Label($"当前方案：{adapter?.Name ?? "未初始化"}");
             GUILayout.Label($"状态：{adapter?.Status ?? "未初始化"}");
-            GUILayout.Label($"对象数量：鱼 {fishCount}，总数 {agents.Count}");
+            GUILayout.Label($"鱼数量：{fishCount} | 对象总数：{agents.Count}");
             GUILayout.Label($"帧率：平均 {metrics.AverageFps:F1} | 1% Low {metrics.OnePercentLowFps:F1}");
-            GUILayout.Label($"模拟耗时：{metrics.SimulationMs:F2} ms | GC：{metrics.GcAllocBytes / 1024f:F1} KB");
+            GUILayout.Label($"模拟耗时：{metrics.SimulationMs:F2} ms | GC {metrics.GcAllocBytes / 1024f:F1} KB");
-            GUILayout.Label($"碰撞调试：XZ重叠 {collisionProbe.HorizontalOverlapPairs} 对 | 3D重叠 {collisionProbe.SpatialOverlapPairs} 对 | 最近间距 {collisionProbe.MinimumClearance:F2}");
+            GUILayout.Label($"速度：极速设定 {(config?.FishSpeed ?? fishMaxSpeed):F1} | 实测峰值 {motionStats.MaxSpeed:F2} | 平均 {motionStats.AverageSpeed:F2}");
-            GUILayout.Label($"帧率限制：不锁帧，vSync={QualitySettings.vSyncCount}，targetFrameRate={Application.targetFrameRate}");
+            GUILayout.Label($"加速度：峰值 {motionStats.MaxAcceleration:F2} | 平均 {motionStats.AverageAcceleration:F2}");
+            GUILayout.Label($"调试倍率：速度 x{speedMultiplier:F2} | 时间 x{timeScale:F2}");
+            GUILayout.Label(hookDropping
+                ? $"鱼钩下落中：{Mathf.Clamp01(hookDropElapsed / Mathf.Max(0.01f, hookDropDuration)) * 100f:F0}%"
+                : "鱼钩状态：未下落");
+            GUILayout.Label($"重叠：XZ {collisionProbe.HorizontalOverlapPairs} | 3D {collisionProbe.SpatialOverlapPairs} | 最近间距 {collisionProbe.MinimumClearance:F2}");
+            GUILayout.Label($"帧率限制：不锁帧，vSync={QualitySettings.vSyncCount}，target={Application.targetFrameRate}");
             GUILayout.Label($"日志样本：{logWriter.SampleCount} | {logWriter.LastExportPath}");
 
+            GUILayout.Space(8);
+            if (GUILayout.Button("投放鱼钩：下落后惊扰争饵", GUILayout.Height(28)))
+            {
+                StartBaitScrambleDemo();
+            }
+
+            GUILayout.Space(8);
+            DrawRuntimeTuningControls();
+
             GUILayout.Space(8);
             DrawEnumButtons("避障方案", framework, value =>
             {
@@ -95,7 +129,7 @@ namespace FishROV.AvoidanceBenchmark
             DrawEnumButtons("观察视角", viewMode, value => viewMode = value);
 
             GUILayout.Space(8);
-            GUILayout.Label($"小鱼数量：{pendingFishCount}");
+            GUILayout.Label($"鱼数量：{pendingFishCount}");
             GUILayout.BeginHorizontal();
             foreach (var preset in CountPresets)
             {
@@ -137,6 +171,7 @@ namespace FishROV.AvoidanceBenchmark
 
         private void OnDestroy()
         {
+            Time.timeScale = 1f;
             logWriter.ExportIfNeeded();
             adapter?.Dispose();
             metrics.Dispose();
@@ -160,6 +195,9 @@ namespace FishROV.AvoidanceBenchmark
                 Scenario = scenario,
                 DimensionMode = dimensionMode,
                 FishCount = fishCount,
+                FishSpeed = scenario == AvoidanceScenarioKind.BaitScramble
+                    ? Mathf.Max(fishMaxSpeed, baitDemoMaxSpeed)
+                    : Mathf.Max(0.1f, fishMaxSpeed),
                 UseSchoolProxy = false
             };
 
@@ -167,12 +205,110 @@ namespace FishROV.AvoidanceBenchmark
             agents.Clear();
             adapter = AvoidanceAdapterRegistry.Create(config.Framework);
             adapter.Initialize(config);
+            BenchmarkAgent.SetSharedConeMaterial(fishMaterial);
 
             SpawnFish();
             scenarioController.Reset(config, agents);
+            if (scenario == AvoidanceScenarioKind.BaitScramble)
+            {
+                DropBaitAtSchoolCenter();
+            }
+            else if (baitMarker != null)
+            {
+                baitMarker.gameObject.SetActive(false);
+            }
+
+            hookDropping = false;
+            motionStats = CalculateMotionStats();
             logWriter.StartRun(config, adapter.Name);
         }
 
+        private void DrawRuntimeTuningControls()
+        {
+            GUILayout.Label($"速度倍率：x{speedMultiplier:F2}");
+            speedMultiplier = GUILayout.HorizontalSlider(speedMultiplier, 0.25f, 4f);
+            GUILayout.BeginHorizontal();
+            if (GUILayout.Button("0.5x"))
+            {
+                speedMultiplier = 0.5f;
+            }
+
+            if (GUILayout.Button("1x"))
+            {
+                speedMultiplier = 1f;
+            }
+
+            if (GUILayout.Button("2x"))
+            {
+                speedMultiplier = 2f;
+            }
+
+            if (GUILayout.Button("4x"))
+            {
+                speedMultiplier = 4f;
+            }
+
+            GUILayout.EndHorizontal();
+
+            GUILayout.Label($"时间缩放：x{timeScale:F2}");
+            var newTimeScale = GUILayout.HorizontalSlider(timeScale, 0.1f, 5f);
+            if (!Mathf.Approximately(newTimeScale, timeScale))
+            {
+                timeScale = newTimeScale;
+                ApplyTimeScale();
+            }
+
+            GUILayout.BeginHorizontal();
+            if (GUILayout.Button("0.25x"))
+            {
+                timeScale = 0.25f;
+                ApplyTimeScale();
+            }
+
+            if (GUILayout.Button("1x"))
+            {
+                timeScale = 1f;
+                ApplyTimeScale();
+            }
+
+            if (GUILayout.Button("2x"))
+            {
+                timeScale = 2f;
+                ApplyTimeScale();
+            }
+
+            if (GUILayout.Button("5x"))
+            {
+                timeScale = 5f;
+                ApplyTimeScale();
+            }
+
+            GUILayout.EndHorizontal();
+        }
+
+        private void ApplyRuntimeSpeed()
+        {
+            if (config == null)
+            {
+                return;
+            }
+
+            var runtimeMaxSpeed = Mathf.Max(0.01f, config.FishSpeed * Mathf.Max(0.01f, speedMultiplier));
+            for (var i = 0; i < agents.Count; i++)
+            {
+                var agent = agents[i];
+                if (agent != null)
+                {
+                    agent.SetMaxSpeed(runtimeMaxSpeed);
+                }
+            }
+        }
+
+        private void ApplyTimeScale()
+        {
+            Time.timeScale = Mathf.Clamp(timeScale, 0.05f, 8f);
+        }
+
         private void SpawnFish()
         {
             for (var i = 0; i < config.FishCount; i++)
@@ -207,6 +343,17 @@ namespace FishROV.AvoidanceBenchmark
                 return new Vector3(Mathf.Cos(angle) * radius, LayerY(index), Mathf.Sin(angle) * radius);
             }
 
+            if (scenario == AvoidanceScenarioKind.BaitScramble)
+            {
+                var angle = index * Mathf.PI * 2f / Mathf.Max(1, total);
+                var band = 0.55f + (index % 7) * 0.055f;
+                var jitter = Random.insideUnitCircle * radius * 0.12f;
+                return new Vector3(
+                    Mathf.Cos(angle) * radius * band + jitter.x,
+                    LayerY(index),
+                    Mathf.Sin(angle) * radius * band + jitter.y);
+            }
+
             var point = Random.insideUnitCircle * radius;
             return new Vector3(point.x, LayerY(index), point.y);
         }
@@ -226,6 +373,77 @@ namespace FishROV.AvoidanceBenchmark
             }
         }
 
+        private void StartBaitScrambleDemo()
+        {
+            scenario = AvoidanceScenarioKind.BaitScramble;
+            StartHookDropAtSchoolCenter();
+        }
+
+        private void StartHookDropAtSchoolCenter()
+        {
+            var position = CalculateAgentCenter();
+            position.y = 0f;
+            hookDropTarget = position;
+            hookDropStart = position + Vector3.up * Mathf.Max(0f, hookDropHeight);
+            hookDropElapsed = 0f;
+            hookDropping = true;
+
+            if (baitMarker != null)
+            {
+                baitMarker.position = hookDropStart;
+                baitMarker.gameObject.SetActive(true);
+            }
+        }
+
+        private void DropBaitAtSchoolCenter()
+        {
+            var position = CalculateAgentCenter();
+            position.y = 0f;
+            DropBaitAtPosition(position);
+        }
+
+        private void DropBaitAtPosition(Vector3 position)
+        {
+            if (config != null)
+            {
+                config.Scenario = AvoidanceScenarioKind.BaitScramble;
+                config.FishSpeed = Mathf.Max(config.FishSpeed, baitDemoMaxSpeed);
+            }
+
+            scenarioController.DropBait(position);
+            if (baitMarker != null)
+            {
+                baitMarker.position = position;
+                baitMarker.gameObject.SetActive(true);
+            }
+        }
+
+        private void UpdateHookDrop(float deltaTime)
+        {
+            if (!hookDropping)
+            {
+                return;
+            }
+
+            hookDropElapsed += deltaTime;
+            var duration = Mathf.Max(0.01f, hookDropDuration);
+            var t = Mathf.Clamp01(hookDropElapsed / duration);
+            var eased = 1f - (1f - t) * (1f - t);
+            var position = Vector3.Lerp(hookDropStart, hookDropTarget, eased);
+
+            if (baitMarker != null)
+            {
+                baitMarker.position = position;
+                baitMarker.gameObject.SetActive(true);
+            }
+
+            if (t >= 1f)
+            {
+                hookDropping = false;
+                DropBaitAtPosition(hookDropTarget);
+            }
+        }
+
         private void EnsureCameraAndLight()
         {
             benchmarkCamera = Camera.main;
@@ -248,6 +466,32 @@ namespace FishROV.AvoidanceBenchmark
             }
         }
 
+        private void EnsureBaitMarker()
+        {
+            if (baitMarker != null)
+            {
+                return;
+            }
+
+            var marker = GameObject.CreatePrimitive(PrimitiveType.Sphere);
+            marker.name = "Benchmark Bait Marker";
+            marker.transform.localScale = Vector3.one * 0.55f;
+            var renderer = marker.GetComponent<Renderer>();
+            if (renderer != null && fishMaterial != null)
+            {
+                renderer.sharedMaterial = fishMaterial;
+            }
+
+            var collider = marker.GetComponent<Collider>();
+            if (collider != null)
+            {
+                collider.enabled = false;
+            }
+
+            baitMarker = marker.transform;
+            marker.SetActive(false);
+        }
+
         private void UpdateCamera()
         {
             if (benchmarkCamera == null || config == null)
@@ -300,6 +544,40 @@ namespace FishROV.AvoidanceBenchmark
             return sum / agents.Count;
         }
 
+        private MotionStats CalculateMotionStats()
+        {
+            if (agents.Count == 0)
+            {
+                return default;
+            }
+
+            var speedSum = 0f;
+            var accelerationSum = 0f;
+            var maxSpeedValue = 0f;
+            var maxAccelerationValue = 0f;
+            for (var i = 0; i < agents.Count; i++)
+            {
+                var agent = agents[i];
+                if (agent == null)
+                {
+                    continue;
+                }
+
+                speedSum += agent.CurrentSpeed;
+                accelerationSum += agent.CurrentAcceleration;
+                maxSpeedValue = Mathf.Max(maxSpeedValue, agent.PeakSpeed);
+                maxAccelerationValue = Mathf.Max(maxAccelerationValue, agent.PeakAcceleration);
+            }
+
+            return new MotionStats
+            {
+                AverageSpeed = speedSum / agents.Count,
+                MaxSpeed = maxSpeedValue,
+                AverageAcceleration = accelerationSum / agents.Count,
+                MaxAcceleration = maxAccelerationValue
+            };
+        }
+
         private static void DrawEnumButtons<T>(string label, T current, System.Action<T> onChanged) where T : System.Enum
         {
             GUILayout.Label(label);
@@ -325,8 +603,10 @@ namespace FishROV.AvoidanceBenchmark
                     return "无避障";
                 case AvoidanceFrameworkKind.Rvo2:
                     return "RVO2";
-                case AvoidanceFrameworkKind.SamplingLocal:
+                case AvoidanceFrameworkKind.AstarProRvo:
-                    return "自研版";
+                    return "A* Pro RVO";
+                case AvoidanceFrameworkKind.AstarProRvoVelocityLike:
+                    return "A* Pro RVO Velocity";
                 case AvoidanceScenarioKind.FreeSwim:
                     return "自由巡游";
                 case AvoidanceScenarioKind.CrossFlow:
@@ -335,6 +615,8 @@ namespace FishROV.AvoidanceBenchmark
                     return "狭窄通道";
                 case AvoidanceScenarioKind.DenseCircleStress:
                     return "圆阵高压";
+                case AvoidanceScenarioKind.BaitScramble:
+                    return "鱼钩争饵";
                 case AvoidanceDimensionMode.XZ2D:
                     return "2D/XZ";
                 case AvoidanceDimensionMode.XZ25D:
@@ -348,10 +630,18 @@ namespace FishROV.AvoidanceBenchmark
                 case BenchmarkViewMode.Side:
                     return "侧视";
                 case BenchmarkViewMode.FollowSchool:
-                    return "跟鱼群";
+                    return "跟随鱼群";
                 default:
                     return value.ToString();
             }
         }
+
+        public struct MotionStats
+        {
+            public float AverageSpeed;
+            public float MaxSpeed;
+            public float AverageAcceleration;
+            public float MaxAcceleration;
+        }
     }
 }

FishROV/Assets/Scripts/Benchmark/FrameworkAdapters/SamplingLocalAvoidanceAdapter.cs

@@ -1,326 +0,0 @@
-using System.Collections.Generic;
-using UnityEngine;
-
-namespace FishROV.AvoidanceBenchmark
-{
-    public sealed class SamplingLocalAvoidanceAdapter : IAvoidanceAdapter
-    {
-        private const float TimeHorizon = 1.25f;
-        private const float NeighbourRangeMultiplier = 5.5f;
-        private const float GoalWeight = 0.85f;
-        private const float SmoothWeight = 0.12f;
-        private const float CollisionWeight = 18f;
-        private const float CurrentOverlapWeight = 80f;
-        private const float BoundaryWeight = 12f;
-        private const int DirectionSamples = 16;
-
-        private readonly Dictionary<BenchmarkAgent, AgentState> states =
-            new Dictionary<BenchmarkAgent, AgentState>(512);
-
-        private readonly List<AgentState> orderedStates = new List<AgentState>(512);
-        private readonly Dictionary<Vector2Int, List<AgentState>> buckets =
-            new Dictionary<Vector2Int, List<AgentState>>(1024);
-
-        private AvoidanceBenchmarkConfig config;
-        private float cellSize;
-
-        public string Name => "采样避障（学习版）";
-        public bool IsAvailable => true;
-        public string Status => "学习版：空间哈希找邻居，采样候选速度，用预测碰撞/平滑/目标偏差/边界惩罚打分。";
-
-        public void Initialize(AvoidanceBenchmarkConfig config)
-        {
-            this.config = config.Clone();
-            states.Clear();
-            orderedStates.Clear();
-            buckets.Clear();
-            cellSize = config.FishRadius * NeighbourRangeMultiplier;
-        }
-
-        public BenchmarkAgent CreateAgent(BenchmarkAgentSpawnInfo spawnInfo, Transform parent)
-        {
-            var agent = BenchmarkAgent.CreatePrimitive(spawnInfo, parent);
-            var state = new AgentState(agent);
-            states[agent] = state;
-            orderedStates.Add(state);
-            return agent;
-        }
-
-        public void SetPreferredVelocity(BenchmarkAgent agent, Vector3 preferredVelocity)
-        {
-            if (!states.TryGetValue(agent, out var state))
-            {
-                agent.ApplyVelocity(preferredVelocity);
-                return;
-            }
-
-            state.PreferredVelocity = ClampForMode(preferredVelocity, agent.MaxSpeed);
-        }
-
-        public void SetTarget(BenchmarkAgent agent, Vector3 target)
-        {
-            agent.Target = target;
-        }
-
-        public void Tick(float deltaTime)
-        {
-            if (deltaTime <= 0f)
-            {
-                return;
-            }
-
-            BuildBuckets();
-
-            for (var i = 0; i < orderedStates.Count; i++)
-            {
-                var state = orderedStates[i];
-                if (state.Agent == null)
-                {
-                    continue;
-                }
-
-                var velocity = ChooseVelocity(state, deltaTime);
-                state.Agent.ApplyVelocity(velocity);
-                state.LastVelocity = velocity;
-            }
-        }
-
-        public void Dispose()
-        {
-            states.Clear();
-            orderedStates.Clear();
-            buckets.Clear();
-        }
-
-        private Vector3 ChooseVelocity(AgentState state, float deltaTime)
-        {
-            var preferred = state.PreferredVelocity;
-            var bestVelocity = preferred;
-            var bestScore = EvaluateVelocity(state, preferred, deltaTime);
-
-            TestCandidate(state, Vector3.zero, deltaTime, ref bestVelocity, ref bestScore);
-
-            var flatPreferred = Flatten(preferred);
-            var baseDirection = flatPreferred.sqrMagnitude > 0.0001f
-                ? flatPreferred.normalized
-                : ForwardFromAgent(state.Agent);
-
-            var maxSpeed = state.Agent.MaxSpeed;
-            for (var i = 0; i < DirectionSamples; i++)
-            {
-                var angle = i * 360f / DirectionSamples;
-                var direction = Quaternion.AngleAxis(angle, Vector3.up) * baseDirection;
-                TestCandidate(state, RestoreVertical(direction * maxSpeed, preferred), deltaTime, ref bestVelocity, ref bestScore);
-                TestCandidate(state, RestoreVertical(direction * maxSpeed * 0.55f, preferred), deltaTime, ref bestVelocity, ref bestScore);
-            }
-
-            return ClampForMode(bestVelocity, maxSpeed);
-        }
-
-        private void TestCandidate(
-            AgentState state,
-            Vector3 candidate,
-            float deltaTime,
-            ref Vector3 bestVelocity,
-            ref float bestScore)
-        {
-            var score = EvaluateVelocity(state, candidate, deltaTime);
-            if (score < bestScore)
-            {
-                bestScore = score;
-                bestVelocity = candidate;
-            }
-        }
-
-        private float EvaluateVelocity(AgentState state, Vector3 candidate, float deltaTime)
-        {
-            candidate = ClampForMode(candidate, state.Agent.MaxSpeed);
-            var preferred = state.PreferredVelocity;
-            var score = GoalWeight * (candidate - preferred).sqrMagnitude;
-            score += SmoothWeight * (candidate - state.LastVelocity).sqrMagnitude;
-            score += BoundaryPenalty(state.Agent.transform.position, candidate, deltaTime);
-
-            var position = state.Agent.transform.position;
-            var centerCell = ToCell(position);
-            var queryRange = Mathf.CeilToInt(
-                state.Agent.Radius * NeighbourRangeMultiplier / cellSize);
-            queryRange = Mathf.Clamp(queryRange, 1, 4);
-
-            for (var x = -queryRange; x <= queryRange; x++)
-            {
-                for (var z = -queryRange; z <= queryRange; z++)
-                {
-                    var cell = new Vector2Int(centerCell.x + x, centerCell.y + z);
-                    if (!buckets.TryGetValue(cell, out var bucket))
-                    {
-                        continue;
-                    }
-
-                    for (var i = 0; i < bucket.Count; i++)
-                    {
-                        var other = bucket[i];
-                        if (other == state || other.Agent == null)
-                        {
-                            continue;
-                        }
-
-                        score += CollisionPenalty(state, other, candidate);
-                    }
-                }
-            }
-
-            return score;
-        }
-
-        private float CollisionPenalty(AgentState state, AgentState other, Vector3 candidate)
-        {
-            var position = state.Agent.transform.position;
-            var otherPosition = other.Agent.transform.position;
-            var relativePosition = otherPosition - position;
-            var combinedRadius = state.Agent.Radius + other.Agent.Radius;
-            var neighbourRange = Mathf.Max(combinedRadius * 2f, state.Agent.Radius * NeighbourRangeMultiplier);
-
-            if (config.DimensionMode != AvoidanceDimensionMode.XYZ3D)
-            {
-                relativePosition.y = 0f;
-            }
-
-            var distanceSq = relativePosition.sqrMagnitude;
-            if (distanceSq > neighbourRange * neighbourRange)
-            {
-                return 0f;
-            }
-
-            var relativeVelocity = candidate - other.LastVelocity;
-            if (config.DimensionMode != AvoidanceDimensionMode.XYZ3D)
-            {
-                relativeVelocity.y = 0f;
-            }
-
-            var clearanceRadius = combinedRadius * 1.08f;
-            var currentDistance = Mathf.Sqrt(Mathf.Max(0.0001f, distanceSq));
-            var currentClearance = currentDistance - clearanceRadius;
-            var penalty = 0f;
-
-            if (currentClearance < 0f)
-            {
-                var overlap = -currentClearance / clearanceRadius;
-                penalty += CurrentOverlapWeight * overlap * overlap;
-            }
-
-            var relativeSpeedSq = relativeVelocity.sqrMagnitude;
-            if (relativeSpeedSq < 0.0001f)
-            {
-                return penalty;
-            }
-
-            var closestTime = Mathf.Clamp(
-                -Vector3.Dot(relativePosition, relativeVelocity) / relativeSpeedSq,
-                0f,
-                TimeHorizon);
-            var closest = relativePosition + relativeVelocity * closestTime;
-            var closestDistance = closest.magnitude;
-            var predictedClearance = closestDistance - clearanceRadius;
-            if (predictedClearance < 0f)
-            {
-                var risk = -predictedClearance / clearanceRadius;
-                var urgency = 1f + (TimeHorizon - closestTime) / TimeHorizon;
-                penalty += CollisionWeight * risk * risk * urgency;
-            }
-
-            return penalty;
-        }
-
-        private float BoundaryPenalty(Vector3 position, Vector3 velocity, float deltaTime)
-        {
-            var next = position + velocity * deltaTime * TimeHorizon;
-            var flat = new Vector2(next.x, next.z);
-            var softRadius = config.ArenaRadius * 0.92f;
-            if (flat.magnitude <= softRadius)
-            {
-                return 0f;
-            }
-
-            var overflow = (flat.magnitude - softRadius) / Mathf.Max(0.001f, config.ArenaRadius - softRadius);
-            return BoundaryWeight * overflow * overflow;
-        }
-
-        private void BuildBuckets()
-        {
-            buckets.Clear();
-            for (var i = 0; i < orderedStates.Count; i++)
-            {
-                var state = orderedStates[i];
-                if (state.Agent == null)
-                {
-                    continue;
-                }
-
-                var cell = ToCell(state.Agent.transform.position);
-                if (!buckets.TryGetValue(cell, out var bucket))
-                {
-                    bucket = new List<AgentState>(8);
-                    buckets[cell] = bucket;
-                }
-
-                bucket.Add(state);
-            }
-        }
-
-        private Vector2Int ToCell(Vector3 position)
-        {
-            return new Vector2Int(
-                Mathf.FloorToInt(position.x / cellSize),
-                Mathf.FloorToInt(position.z / cellSize));
-        }
-
-        private Vector3 ClampForMode(Vector3 velocity, float maxSpeed)
-        {
-            if (config.DimensionMode != AvoidanceDimensionMode.XYZ3D)
-            {
-                velocity.y = 0f;
-            }
-
-            return Vector3.ClampMagnitude(velocity, maxSpeed);
-        }
-
-        private Vector3 RestoreVertical(Vector3 flatVelocity, Vector3 preferred)
-        {
-            if (config.DimensionMode == AvoidanceDimensionMode.XYZ3D)
-            {
-                flatVelocity.y = preferred.y;
-            }
-            else
-            {
-                flatVelocity.y = 0f;
-            }
-
-            return flatVelocity;
-        }
-
-        private static Vector3 Flatten(Vector3 value)
-        {
-            value.y = 0f;
-            return value;
-        }
-
-        private static Vector3 ForwardFromAgent(BenchmarkAgent agent)
-        {
-            var forward = agent != null ? agent.transform.forward : Vector3.forward;
-            forward.y = 0f;
-            return forward.sqrMagnitude > 0.0001f ? forward.normalized : Vector3.forward;
-        }
-
-        private sealed class AgentState
-        {
-            public AgentState(BenchmarkAgent agent)
-            {
-                Agent = agent;
-            }
-
-            public BenchmarkAgent Agent { get; }
-            public Vector3 PreferredVelocity { get; set; }
-            public Vector3 LastVelocity { get; set; }
-        }
-    }
-}

FishROV/Assets/Scripts/Benchmark/FrameworkAdapters/SamplingLocalAvoidanceAdapter.cs.meta

@@ -1,11 +0,0 @@
-fileFormatVersion: 2
-guid: 41e3ccd900cb4a37bf60c67fa49c906c
-MonoImporter:
-  externalObjects: {}
-  serializedVersion: 2
-  defaultReferences: []
-  executionOrder: 0
-  icon: {instanceID: 0}
-  userData: 
-  assetBundleName: 
-  assetBundleVariant: 

FishROV/Assets/Scripts/Benchmark/Scenarios/AvoidanceScenarioController.cs

@@ -8,11 +8,15 @@ namespace FishROV.AvoidanceBenchmark
         private readonly Dictionary<BenchmarkAgent, Vector3> targets = new Dictionary<BenchmarkAgent, Vector3>();
         private AvoidanceBenchmarkConfig config;
         private float elapsed;
+        private float baitDropTime;
+        private Vector3 baitPosition;
 
         public void Reset(AvoidanceBenchmarkConfig benchmarkConfig, IReadOnlyList<BenchmarkAgent> agents)
         {
             config = benchmarkConfig.Clone();
             elapsed = 0f;
+            baitDropTime = 0f;
+            baitPosition = Vector3.zero;
             targets.Clear();
 
             foreach (var agent in agents)
@@ -21,16 +25,34 @@ namespace FishROV.AvoidanceBenchmark
             }
         }
 
-        public Vector3 GetPreferredVelocity(BenchmarkAgent agent, float deltaTime)
+        public void Tick(float deltaTime)
         {
             elapsed += deltaTime;
+        }
 
+        public void DropBait(Vector3 position)
+        {
+            baitPosition = position;
+            baitDropTime = elapsed;
+            if (config != null)
+            {
+                config.Scenario = AvoidanceScenarioKind.BaitScramble;
+            }
+        }
+
+        public Vector3 GetPreferredVelocity(BenchmarkAgent agent, float deltaTime)
+        {
             if (!targets.TryGetValue(agent, out var target))
             {
                 target = CreateTarget(agent);
                 targets[agent] = target;
             }
 
+            if (config.Scenario == AvoidanceScenarioKind.BaitScramble)
+            {
+                return GetBaitScrambleVelocity(agent);
+            }
+
             if ((target - agent.transform.position).sqrMagnitude < 4f)
             {
                 target = CreateTarget(agent);
@@ -57,6 +79,8 @@ namespace FishROV.AvoidanceBenchmark
 
             switch (config.Scenario)
             {
+                case AvoidanceScenarioKind.BaitScramble:
+                    return baitPosition;
                 case AvoidanceScenarioKind.CrossFlow:
                     return agent.GroupId % 2 == 0
                         ? new Vector3(radius, HeightFor(agent), Random.Range(-radius, radius))
@@ -87,5 +111,47 @@ namespace FishROV.AvoidanceBenchmark
 
             return Random.Range(-4f, 4f);
         }
+
+        private Vector3 GetBaitScrambleVelocity(BenchmarkAgent agent)
+        {
+            var position = agent.transform.position;
+            var flatOffset = new Vector3(position.x - baitPosition.x, 0f, position.z - baitPosition.z);
+            if (flatOffset.sqrMagnitude < 0.0001f)
+            {
+                flatOffset = Quaternion.Euler(0f, agent.GroupId * 137.5f, 0f) * Vector3.forward;
+            }
+
+            var outward = flatOffset.normalized;
+            var tangentSign = agent.GroupId % 2 == 0 ? 1f : -1f;
+            var tangent = new Vector3(-outward.z, 0f, outward.x) * tangentSign;
+            var distance = flatOffset.magnitude;
+            var age = elapsed - baitDropTime;
+
+            Vector3 desired;
+            if (age < 1.2f)
+            {
+                desired = outward * agent.MaxSpeed * 1.15f + tangent * agent.MaxSpeed * 0.25f;
+            }
+            else
+            {
+                var ringRadius = Mathf.Lerp(5.5f, 2.2f, Mathf.Clamp01((age - 1.2f) / 7.5f));
+                var radialError = distance - ringRadius;
+                var radial = -outward * Mathf.Clamp(radialError * 0.55f, -0.75f, 1f);
+                var swirl = tangent * Mathf.Lerp(0.35f, 0.95f, Mathf.Clamp01((age - 1.2f) / 5f));
+                var baitPull = -outward * Mathf.Lerp(0.25f, 0.55f, Mathf.Clamp01((age - 1.2f) / 4f));
+                desired = (radial + swirl + baitPull).normalized * agent.MaxSpeed;
+            }
+
+            if (config.DimensionMode == AvoidanceDimensionMode.XYZ3D)
+            {
+                desired.y = Mathf.Clamp(baitPosition.y - position.y, -1f, 1f) * agent.MaxSpeed * 0.35f;
+            }
+            else
+            {
+                desired.y = 0f;
+            }
+
+            return desired;
+        }
     }
 }

FishROV/Docs/avoidance-benchmark-report-20260613.html

@@ -0,0 +1,637 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <meta name="color-scheme" content="light">
+  <title>FishROV RVO2 本地避障 Benchmark 报告</title>
+  <style>
+    :root {
+      --ink: #172033;
+      --muted: #667085;
+      --paper: #ffffff;
+      --canvas: #eef2f8;
+      --line: #dce3ee;
+      --navy: #102a43;
+      --blue: #2f6fed;
+      --blue-soft: #eaf1ff;
+      --cyan: #13a8a8;
+      --cyan-soft: #e6fafa;
+      --green: #138a5b;
+      --green-soft: #e7f7ef;
+      --orange: #e77817;
+      --orange-soft: #fff2e5;
+      --red: #c73e4d;
+      --red-soft: #ffeaed;
+      --gold: #c88a10;
+      --gold-soft: #fff7d8;
+      --shadow: 0 14px 40px rgba(16, 42, 67, .09);
+    }
+
+    * { box-sizing: border-box; }
+    html { scroll-behavior: smooth; }
+
+    body {
+      margin: 0;
+      color: var(--ink);
+      background:
+        radial-gradient(circle at 8% 0%, rgba(47, 111, 237, .11), transparent 28rem),
+        linear-gradient(180deg, #f7f9fc 0, var(--canvas) 100%);
+      font: 14px/1.7 -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC", "Microsoft YaHei", sans-serif;
+    }
+
+    a { color: inherit; }
+
+    .shell {
+      display: grid;
+      grid-template-columns: 248px minmax(0, 1180px);
+      gap: 26px;
+      width: min(1480px, calc(100% - 32px));
+      margin: 24px auto 64px;
+      align-items: start;
+    }
+
+    .side {
+      position: sticky;
+      top: 20px;
+      min-width: 0;
+      padding: 20px 16px;
+      border: 1px solid var(--line);
+      border-radius: 16px;
+      background: rgba(255, 255, 255, .94);
+      box-shadow: var(--shadow);
+      backdrop-filter: blur(10px);
+    }
+
+    .brand {
+      display: flex;
+      gap: 11px;
+      align-items: center;
+      padding: 2px 5px 17px;
+      border-bottom: 1px solid var(--line);
+    }
+
+    .brand-mark {
+      display: grid;
+      width: 38px;
+      height: 38px;
+      place-items: center;
+      border-radius: 11px;
+      color: #fff;
+      background: linear-gradient(145deg, var(--blue), #1744ad);
+      font-size: 18px;
+      font-weight: 800;
+      box-shadow: 0 8px 16px rgba(47, 111, 237, .28);
+    }
+
+    .brand strong { display: block; font-size: 15px; }
+    .brand span { display: block; color: var(--muted); font-size: 11px; }
+
+    .side nav { margin-top: 14px; }
+    .side nav a {
+      display: block;
+      padding: 7px 10px;
+      border-radius: 8px;
+      color: #46536a;
+      text-decoration: none;
+      font-size: 13px;
+    }
+    .side nav a:hover {
+      color: var(--blue);
+      background: var(--blue-soft);
+    }
+
+    .side-note {
+      margin: 15px 5px 0;
+      padding-top: 14px;
+      border-top: 1px solid var(--line);
+      color: var(--muted);
+      font-size: 11px;
+    }
+
+    main { min-width: 0; }
+
+    .hero, .section {
+      border: 1px solid var(--line);
+      border-radius: 18px;
+      background: var(--paper);
+      box-shadow: var(--shadow);
+    }
+
+    .hero {
+      position: relative;
+      overflow: hidden;
+      padding: 38px 40px 34px;
+      color: #fff;
+      background:
+        linear-gradient(115deg, rgba(16, 42, 67, .98), rgba(23, 64, 112, .96) 55%, rgba(47, 111, 237, .92)),
+        #102a43;
+    }
+
+    .hero::after {
+      position: absolute;
+      right: -55px;
+      bottom: -115px;
+      width: 330px;
+      height: 330px;
+      border: 45px solid rgba(255, 255, 255, .055);
+      border-radius: 50%;
+      content: "";
+    }
+
+    .eyebrow {
+      margin: 0 0 10px;
+      color: #a8c8ff;
+      font-size: 12px;
+      font-weight: 750;
+      letter-spacing: .13em;
+      text-transform: uppercase;
+    }
+
+    h1 {
+      max-width: 850px;
+      margin: 0;
+      font-size: clamp(30px, 5vw, 48px);
+      line-height: 1.18;
+      letter-spacing: -.025em;
+    }
+
+    .lead {
+      max-width: 850px;
+      margin: 16px 0 0;
+      color: #d8e6fb;
+      font-size: 16px;
+    }
+
+    .hero-meta, .summary-grid, .grid-2, .grid-3 {
+      display: grid;
+      gap: 14px;
+    }
+
+    .hero-meta {
+      grid-template-columns: repeat(4, minmax(0, 1fr));
+      margin-top: 28px;
+    }
+
+    .hero-meta div {
+      padding: 12px 14px;
+      border: 1px solid rgba(255, 255, 255, .13);
+      border-radius: 11px;
+      background: rgba(255, 255, 255, .07);
+    }
+
+    .hero-meta span { display: block; color: #a9bdd8; font-size: 11px; }
+    .hero-meta strong { display: block; margin-top: 2px; font-size: 14px; }
+
+    .section {
+      margin-top: 20px;
+      padding: 28px 30px;
+      scroll-margin-top: 24px;
+    }
+
+    .section-head {
+      display: flex;
+      gap: 20px;
+      align-items: flex-start;
+      justify-content: space-between;
+      margin-bottom: 20px;
+    }
+
+    .kicker {
+      margin: 0 0 3px;
+      color: var(--blue);
+      font-size: 11px;
+      font-weight: 800;
+      letter-spacing: .12em;
+      text-transform: uppercase;
+    }
+
+    h2 { margin: 0; color: var(--navy); font-size: 24px; line-height: 1.3; }
+    h3 { margin: 0 0 10px; color: var(--navy); font-size: 16px; }
+    p { margin: 8px 0; }
+    .muted { color: var(--muted); }
+
+    .summary-grid { grid-template-columns: repeat(4, minmax(0, 1fr)); }
+    .metric {
+      padding: 17px;
+      border: 1px solid var(--line);
+      border-radius: 13px;
+      background: #fbfcfe;
+    }
+    .metric b { display: block; color: var(--blue); font-size: 27px; line-height: 1.2; }
+    .metric span { display: block; margin-top: 6px; color: var(--muted); font-size: 12px; }
+
+    .grid-2 { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+    .grid-3 { grid-template-columns: repeat(3, minmax(0, 1fr)); }
+    .grid-2 > *, .grid-3 > * { min-width: 0; }
+
+    .info-card {
+      padding: 17px 18px;
+      border: 1px solid var(--line);
+      border-radius: 13px;
+      background: #fbfcfe;
+    }
+    .info-card p:last-child { margin-bottom: 0; }
+
+    .tag {
+      display: inline-block;
+      margin: 2px 5px 2px 0;
+      padding: 3px 9px;
+      border-radius: 999px;
+      color: #31507d;
+      background: var(--blue-soft);
+      font-size: 11px;
+      font-weight: 700;
+    }
+    .tag.green { color: var(--green); background: var(--green-soft); }
+    .tag.orange { color: #9a4e05; background: var(--orange-soft); }
+    .tag.red { color: var(--red); background: var(--red-soft); }
+    .tag.cyan { color: #087273; background: var(--cyan-soft); }
+
+    .callout {
+      margin: 17px 0 0;
+      padding: 14px 16px;
+      border-left: 4px solid var(--blue);
+      border-radius: 9px;
+      background: var(--blue-soft);
+    }
+    .callout.warn { border-color: var(--orange); background: var(--orange-soft); }
+    .callout.good { border-color: var(--green); background: var(--green-soft); }
+
+    .compare-card {
+      padding: 17px;
+      border: 1px solid var(--line);
+      border-radius: 13px;
+      background: #fbfcfe;
+    }
+    .compare-card .rule {
+      margin-bottom: 13px;
+      padding-bottom: 11px;
+      border-bottom: 1px solid var(--line);
+      color: var(--muted);
+      font-size: 12px;
+    }
+    .compare-card .rule strong { color: var(--navy); }
+
+    table {
+      width: 100%;
+      border-collapse: collapse;
+      font-size: 13px;
+    }
+    th, td {
+      padding: 10px 9px;
+      border-bottom: 1px solid var(--line);
+      text-align: right;
+      vertical-align: top;
+      white-space: nowrap;
+      font-variant-numeric: tabular-nums;
+    }
+    th {
+      color: #31507d;
+      background: #f3f7fd;
+      font-size: 12px;
+      font-weight: 800;
+    }
+    th:first-child, td:first-child,
+    th:nth-child(2), td:nth-child(2) { text-align: left; }
+    .wide-table { overflow-x: auto; }
+
+    .score {
+      display: inline-flex;
+      min-width: 68px;
+      justify-content: center;
+      padding: 2px 8px;
+      border-radius: 999px;
+      font-size: 12px;
+      font-weight: 800;
+    }
+    .score.good { color: var(--green); background: var(--green-soft); }
+    .score.warn { color: #9a4e05; background: var(--orange-soft); }
+    .score.bad { color: var(--red); background: var(--red-soft); }
+    .good-text { color: var(--green); font-weight: 800; }
+    .warn-text { color: var(--orange); font-weight: 800; }
+    .bad-text { color: var(--red); font-weight: 800; }
+
+    .bar-list {
+      display: grid;
+      gap: 10px;
+      margin-top: 12px;
+    }
+    .bar-row {
+      display: grid;
+      grid-template-columns: 106px 1fr 68px;
+      gap: 10px;
+      align-items: center;
+      font-size: 12px;
+      font-weight: 750;
+      font-variant-numeric: tabular-nums;
+    }
+    .track {
+      height: 14px;
+      border-radius: 999px;
+      background: #e8eef7;
+      overflow: hidden;
+    }
+    .fill {
+      height: 100%;
+      border-radius: 999px;
+      background: var(--blue);
+    }
+    .fill.green { background: var(--green); }
+    .fill.orange { background: var(--orange); }
+    .fill.red { background: var(--red); }
+
+    .foot {
+      margin-top: 18px;
+      color: var(--muted);
+      font-size: 12px;
+    }
+
+    @media (max-width: 1080px) {
+      .shell { grid-template-columns: 1fr; }
+      .side { position: static; }
+      .side nav { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); }
+      .hero-meta, .summary-grid, .grid-2, .grid-3 { grid-template-columns: 1fr; }
+      .section-head { display: block; }
+    }
+
+    @media (max-width: 640px) {
+      .shell { width: min(100% - 20px, 1480px); margin-top: 10px; }
+      .hero, .section { border-radius: 14px; }
+      .hero { padding: 28px 22px; }
+      .section { padding: 22px 18px; }
+      .side nav { grid-template-columns: 1fr; }
+      .bar-row { grid-template-columns: 86px 1fr 56px; }
+    }
+  </style>
+</head>
+<body>
+  <div class="shell">
+    <aside class="side">
+      <div class="brand">
+        <div class="brand-mark">FR</div>
+        <div>
+          <strong>FishROV Benchmark</strong>
+          <span>RVO2 local avoidance</span>
+        </div>
+      </div>
+      <nav>
+        <a href="#summary">结论摘要</a>
+        <a href="#phone">百元机内对比</a>
+        <a href="#editor">编辑器内对比</a>
+        <a href="#cross">编辑器 vs 百元机</a>
+        <a href="#method">控制变量口径</a>
+        <a href="#actions">建议动作</a>
+      </nav>
+      <div class="side-note">
+        统计口径：排除 timeSeconds = 0 的启动采样。比较表均标注“固定条件”和“唯一变量”。
+      </div>
+    </aside>
+
+    <main>
+      <section class="hero">
+        <p class="eyebrow">2026-06-13 / AvoidanceBenchmarkLogs</p>
+        <h1>RVO2 本地避障结果：按唯一变量拆开看</h1>
+        <p class="lead">这版报告不再把编辑器和百元机混成一张表。先在百元机内部比较算法、数量、维度；再在编辑器内部看同样趋势；最后才做同配置的跨环境差异。</p>
+        <div class="hero-meta">
+          <div><span>移动端主依据</span><strong>百元机 / 720x1600</strong></div>
+          <div><span>趋势参考</span><strong>编辑器 / 750x617</strong></div>
+          <div><span>推荐默认档</span><strong>300 鱼 + XYZ3D</strong></div>
+          <div><span>压测红线</span><strong>500 鱼开始警戒</strong></div>
+        </div>
+      </section>
+
+      <section id="summary" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Executive Summary</p>
+            <h2>结论摘要</h2>
+          </div>
+          <p class="muted">先给结论，再看证据。所有跨表比较都避免同时改变多个核心变量。</p>
+        </div>
+        <div class="summary-grid">
+          <div class="metric"><b>300</b><span>百元机可用档。RVO2 / FreeSwim / XZ25D 第二次运行达到 30.1 FPS，空间重叠均值 1.1。</span></div>
+          <div class="metric"><b>500</b><span>百元机压测线。XYZ3D 在 FreeSwim 仍有 26.2 FPS，但 1% Low 已降到 15.8。</span></div>
+          <div class="metric"><b>XYZ3D</b><span>同场景同数量下空间重叠最低，Dense/CrossFlow 1000 都明显优于 XZ2D/XZ25D。</span></div>
+          <div class="metric"><b>6-14x</b><span>同配置下，编辑器 FPS 约为百元机 6 倍左右；百元机 simulation ms 约为编辑器 10-14 倍。</span></div>
+        </div>
+        <div class="callout good">
+          <strong>推荐口径：</strong>玩法默认目标按百元机结果定，不按编辑器定。编辑器只用于观察算法随数量和维度变化的趋势。
+        </div>
+      </section>
+
+      <section id="phone" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Phone Only / Controlled Variables</p>
+            <h2>百元机内对比</h2>
+          </div>
+          <p class="muted">同一设备环境内比较，避免把硬件差异混进算法结论。</p>
+        </div>
+
+        <div class="grid-2">
+          <article class="compare-card">
+            <h3>算法变量：无避障 vs RVO2</h3>
+            <div class="rule"><strong>固定：</strong>百元机 / FreeSwim / XZ25D / 300 鱼。<strong>唯一变量：</strong>避障策略。</div>
+            <table>
+              <thead><tr><th>策略</th><th>FPS</th><th>1% Low</th><th>空间重叠</th><th>最小间距</th><th>判断</th></tr></thead>
+              <tbody>
+                <tr><td>无避障</td><td>7.0</td><td>0.5</td><td>80.9</td><td>-0.674</td><td><span class="score bad">不可用</span></td></tr>
+                <tr><td>RVO2 第一次</td><td>26.4</td><td>14.4</td><td>1.6</td><td>-0.298</td><td><span class="score warn">有尖峰</span></td></tr>
+                <tr><td>RVO2 第二次</td><td>30.1</td><td>23.6</td><td>1.1</td><td>-0.017</td><td><span class="score good">可用</span></td></tr>
+              </tbody>
+            </table>
+            <p class="muted">RVO2 不是只减少重叠，也显著改善了百元机上的可运行性。第一次与第二次 RVO2 结果不同，后续建议保留重复运行取中位数。</p>
+          </article>
+
+          <article class="compare-card">
+            <h3>数量变量：FreeSwim / XYZ3D</h3>
+            <div class="rule"><strong>固定：</strong>百元机 / RVO2 / FreeSwim / XYZ3D。<strong>唯一变量：</strong>鱼数量。</div>
+            <div class="bar-list">
+              <div class="bar-row"><span>100 鱼</span><div class="track"><div class="fill green" style="width:100%"></div></div><strong>30.2 FPS</strong></div>
+              <div class="bar-row"><span>500 鱼</span><div class="track"><div class="fill orange" style="width:86.8%"></div></div><strong>26.2 FPS</strong></div>
+              <div class="bar-row"><span>800 鱼</span><div class="track"><div class="fill red" style="width:39.1%"></div></div><strong>11.8 FPS</strong></div>
+              <div class="bar-row"><span>1500 鱼</span><div class="track"><div class="fill red" style="width:29.5%"></div></div><strong>8.9 FPS</strong></div>
+            </div>
+            <table>
+              <thead><tr><th>数量</th><th>FPS</th><th>1% Low</th><th>Sim ms</th><th>P95 ms</th><th>空间重叠</th></tr></thead>
+              <tbody>
+                <tr><td>100</td><td>30.2</td><td>28.6</td><td>4.43</td><td>7.82</td><td>0.0</td></tr>
+                <tr><td>500</td><td>26.2</td><td>15.8</td><td>22.70</td><td>37.08</td><td>0.7</td></tr>
+                <tr><td>800</td><td>11.8</td><td>5.2</td><td>54.07</td><td>104.84</td><td>7.7</td></tr>
+                <tr><td>1500</td><td>8.9</td><td>3.1</td><td>103.64</td><td>189.01</td><td>84.2</td></tr>
+              </tbody>
+            </table>
+          </article>
+        </div>
+
+        <div class="grid-2" style="margin-top:14px;">
+          <article class="compare-card">
+            <h3>维度变量：Dense 1000</h3>
+            <div class="rule"><strong>固定：</strong>百元机 / RVO2 / DenseCircleStress / 1000 鱼。<strong>唯一变量：</strong>维度。</div>
+            <table>
+              <thead><tr><th>维度</th><th>FPS</th><th>Sim ms</th><th>P95 ms</th><th>空间重叠</th><th>判断</th></tr></thead>
+              <tbody>
+                <tr><td>XYZ3D</td><td>14.2</td><td>55.33</td><td>103.78</td><td class="good-text">87.1</td><td><span class="score warn">质量最好</span></td></tr>
+                <tr><td>XZ25D</td><td>12.1</td><td>69.90</td><td>144.30</td><td class="warn-text">474.4</td><td><span class="score bad">偏高</span></td></tr>
+                <tr><td>XZ2D</td><td>11.8</td><td>70.63</td><td>131.34</td><td class="bad-text">1190.3</td><td><span class="score bad">最差</span></td></tr>
+              </tbody>
+            </table>
+          </article>
+
+          <article class="compare-card">
+            <h3>维度变量：CrossFlow 1000</h3>
+            <div class="rule"><strong>固定：</strong>百元机 / RVO2 / CrossFlow / 1000 鱼。<strong>唯一变量：</strong>维度。</div>
+            <table>
+              <thead><tr><th>维度</th><th>FPS</th><th>Sim ms</th><th>P95 ms</th><th>空间重叠</th><th>判断</th></tr></thead>
+              <tbody>
+                <tr><td>XYZ3D</td><td>13.1</td><td>60.92</td><td>117.81</td><td class="good-text">38.5</td><td><span class="score warn">质量最好</span></td></tr>
+                <tr><td>XZ25D</td><td>11.9</td><td>77.62</td><td>202.03</td><td class="warn-text">384.4</td><td><span class="score bad">尾延迟高</span></td></tr>
+                <tr><td>XZ2D</td><td>12.2</td><td>60.66</td><td>118.03</td><td class="bad-text">714.1</td><td><span class="score bad">重叠高</span></td></tr>
+              </tbody>
+            </table>
+          </article>
+        </div>
+      </section>
+
+      <section id="editor" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Editor Only / Controlled Variables</p>
+            <h2>编辑器内对比</h2>
+          </div>
+          <p class="muted">编辑器不用于移动端性能结论，但可以看数量增长和维度变化的趋势是否一致。</p>
+        </div>
+
+        <div class="grid-2">
+          <article class="compare-card">
+            <h3>数量变量：FreeSwim / XYZ3D</h3>
+            <div class="rule"><strong>固定：</strong>编辑器 / RVO2 / FreeSwim / XYZ3D。<strong>唯一变量：</strong>鱼数量。</div>
+            <table>
+              <thead><tr><th>数量</th><th>FPS</th><th>1% Low</th><th>Sim ms</th><th>P95 ms</th><th>空间重叠</th></tr></thead>
+              <tbody>
+                <tr><td>100</td><td>169.4</td><td>93.4</td><td>0.62</td><td>0.84</td><td>0.0</td></tr>
+                <tr><td>300</td><td>134.0</td><td>72.7</td><td>1.53</td><td>2.17</td><td>0.0</td></tr>
+                <tr><td>500</td><td>129.7</td><td>67.8</td><td>2.33</td><td>2.41</td><td>0.2</td></tr>
+                <tr><td>1000</td><td>71.5</td><td>33.0</td><td>4.73</td><td>4.98</td><td>5.1</td></tr>
+                <tr><td>2000</td><td>40.5</td><td>17.7</td><td>12.18</td><td>25.97</td><td>55.3</td></tr>
+                <tr><td>5000</td><td>18.2</td><td>5.7</td><td>36.22</td><td>86.95</td><td>774.7</td></tr>
+              </tbody>
+            </table>
+          </article>
+
+          <article class="compare-card">
+            <h3>维度变量：Dense 1000 / 2000</h3>
+            <div class="rule"><strong>固定：</strong>编辑器 / RVO2 / DenseCircleStress / 同数量。<strong>唯一变量：</strong>维度。</div>
+            <table>
+              <thead><tr><th>数量</th><th>维度</th><th>FPS</th><th>Sim ms</th><th>P95 ms</th><th>空间重叠</th></tr></thead>
+              <tbody>
+                <tr><td>1000</td><td>XYZ3D</td><td>85.2</td><td>5.38</td><td>8.21</td><td class="good-text">29.5</td></tr>
+                <tr><td>1000</td><td>XZ25D</td><td>83.8</td><td>4.97</td><td>5.43</td><td class="warn-text">264.6</td></tr>
+                <tr><td>1000</td><td>XZ2D</td><td>79.8</td><td>5.41</td><td>7.25</td><td class="bad-text">851.9</td></tr>
+                <tr><td>2000</td><td>XYZ3D</td><td>55.1</td><td>11.14</td><td>26.22</td><td class="good-text">154.2</td></tr>
+                <tr><td>2000</td><td>XZ25D</td><td>50.2</td><td>12.17</td><td>26.58</td><td class="warn-text">765.8</td></tr>
+                <tr><td>2000</td><td>XZ2D</td><td>48.7</td><td>10.34</td><td>11.58</td><td class="bad-text">2093.6</td></tr>
+              </tbody>
+            </table>
+          </article>
+        </div>
+
+        <div class="callout">
+          <strong>编辑器趋势与百元机一致：</strong>XYZ3D 的空间重叠始终最低；XZ2D 在 Dense 场景中最容易堆叠。不同的是，编辑器性能余量会掩盖百元机上的真实性能红线。
+        </div>
+      </section>
+
+      <section id="cross" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Cross Environment / Same Config</p>
+            <h2>编辑器 vs 百元机</h2>
+          </div>
+          <p class="muted">同配置跨环境对比只能说明设备差异，不用于证明算法优劣。</p>
+        </div>
+
+        <div class="wide-table compare-card">
+          <div class="rule"><strong>固定：</strong>RVO2 / 同场景 / 同维度 / 同数量。<strong>唯一变量：</strong>运行环境。</div>
+          <table>
+            <thead>
+              <tr>
+                <th>配置</th>
+                <th>编辑器 FPS</th>
+                <th>百元机 FPS</th>
+                <th>FPS 倍率</th>
+                <th>编辑器 Sim ms</th>
+                <th>百元机 Sim ms</th>
+                <th>Sim 倍率</th>
+                <th>空间重叠：编辑器 / 百元机</th>
+              </tr>
+            </thead>
+            <tbody>
+              <tr><td>FreeSwim / XYZ3D / 100</td><td>169.4</td><td>30.2</td><td>5.6x</td><td>0.62</td><td>4.43</td><td>7.1x</td><td>0.0 / 0.0</td></tr>
+              <tr><td>FreeSwim / XYZ3D / 500</td><td>129.7</td><td>26.2</td><td>5.0x</td><td>2.33</td><td>22.70</td><td>9.7x</td><td>0.2 / 0.7</td></tr>
+              <tr><td>Dense / XYZ3D / 1000</td><td>85.2</td><td>14.2</td><td>6.0x</td><td>5.38</td><td>55.33</td><td>10.3x</td><td>29.5 / 87.1</td></tr>
+              <tr><td>Dense / XZ25D / 1000</td><td>83.8</td><td>12.1</td><td>6.9x</td><td>4.97</td><td>69.90</td><td>14.1x</td><td>264.6 / 474.4</td></tr>
+              <tr><td>Dense / XZ2D / 1000</td><td>79.8</td><td>11.8</td><td>6.8x</td><td>5.41</td><td>70.63</td><td>13.1x</td><td>851.9 / 1190.3</td></tr>
+            </tbody>
+          </table>
+        </div>
+
+        <div class="callout warn">
+          <strong>读法：</strong>跨环境对比的唯一变量是“运行环境”，因此它只回答“编辑器结果离百元机有多远”。它不能替代百元机内的算法、数量、维度对比。
+        </div>
+      </section>
+
+      <section id="method" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Method</p>
+            <h2>控制变量口径</h2>
+          </div>
+        </div>
+
+        <div class="grid-3">
+          <div class="info-card">
+            <h3>百元机内对比</h3>
+            <p><span class="tag green">环境固定</span><span class="tag">百元机</span></p>
+            <p class="muted">用于回答移动端真实可用性：算法是否有效、数量红线在哪里、哪个维度质量最好。</p>
+          </div>
+          <div class="info-card">
+            <h3>编辑器内对比</h3>
+            <p><span class="tag cyan">环境固定</span><span class="tag">编辑器</span></p>
+            <p class="muted">用于观察趋势，不直接判断移动端上线规模。</p>
+          </div>
+          <div class="info-card">
+            <h3>跨环境对比</h3>
+            <p><span class="tag orange">配置固定</span><span class="tag">环境变化</span></p>
+            <p class="muted">用于量化编辑器到百元机的性能折损，避免把桌面余量误当移动端余量。</p>
+          </div>
+        </div>
+      </section>
+
+      <section id="actions" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Next Actions</p>
+            <h2>建议动作</h2>
+          </div>
+        </div>
+        <div class="grid-3">
+          <div class="info-card">
+            <h3>玩法默认档</h3>
+            <p>按百元机结果，把默认目标定为 <strong>300 鱼 + XYZ3D</strong>。500 鱼作为压测档，不作为默认玩法规模。</p>
+          </div>
+          <div class="info-card">
+            <h3>优化方向</h3>
+            <p>500+ 鱼需要分级避障：远场降频或关闭 RVO，只对近场、玩家交互区和高风险鱼群启用。</p>
+          </div>
+          <div class="info-card">
+            <h3>后续 Benchmark</h3>
+            <p>补齐设备、构建类型、VSync、targetFrameRate、Burst/Jobs 状态，并对同配置至少跑 3 次取中位数。</p>
+          </div>
+        </div>
+        <p class="foot">数据源：编辑器批次来自 C:\Users\13999\AppData\LocalLow\DefaultCompany\FishROV\AvoidanceBenchmarkLogs；百元机批次来自 C:\Users\13999\AppData\Local\Temp。</p>
+      </section>
+    </main>
+  </div>
+</body>
+</html>

FishROV/Docs/avoidance-benchmark-report-20260615-android2.html

@@ -0,0 +1,153 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+<meta charset="utf-8">
+<meta name="viewport" content="width=device-width, initial-scale=1">
+<title>FishROV Android 2 避障复测报告</title>
+<style>
+:root{--ink:#172033;--muted:#667085;--paper:#fff;--canvas:#eef2f8;--line:#dce3ee;--blue:#2f6fed;--blue-soft:#eaf1ff;--green:#138a5b;--green-soft:#e7f7ef;--orange:#d96b11;--orange-soft:#fff2e5;--red:#c73e4d;--shadow:0 14px 40px rgba(16,42,67,.09)}
+*{box-sizing:border-box}body{margin:0;color:var(--ink);background:linear-gradient(180deg,#f7f9fc 0,var(--canvas) 100%);font:14px/1.65 -apple-system,BlinkMacSystemFont,"Segoe UI","PingFang SC","Microsoft YaHei",sans-serif}.shell{display:grid;grid-template-columns:248px minmax(0,1180px);gap:26px;width:min(1480px,calc(100% - 32px));margin:24px auto 64px;align-items:start}.side{position:sticky;top:20px;padding:20px 16px;border:1px solid var(--line);border-radius:16px;background:rgba(255,255,255,.94);box-shadow:var(--shadow)}.brand{display:flex;gap:11px;align-items:center;padding:2px 5px 17px;border-bottom:1px solid var(--line)}.brand-mark{display:grid;width:38px;height:38px;place-items:center;border-radius:11px;color:#fff;background:linear-gradient(145deg,var(--blue),#1744ad);font-weight:800}.brand strong{display:block;font-size:15px}.brand span{color:var(--muted);font-size:11px}nav{margin-top:14px}nav a{display:block;padding:7px 10px;border-radius:8px;color:#46536a;text-decoration:none;font-size:13px}nav a:hover{color:var(--blue);background:var(--blue-soft)}main{min-width:0}.hero,.section{border:1px solid var(--line);border-radius:18px;background:var(--paper);box-shadow:var(--shadow)}.hero{padding:38px 40px 34px;color:#fff;background:linear-gradient(115deg,rgba(16,42,67,.98),rgba(23,64,112,.96) 56%,rgba(47,111,237,.92))}.eyebrow{margin:0 0 8px;color:#cfe1ff;font-weight:700;letter-spacing:.02em;text-transform:uppercase;font-size:12px}h1{margin:0;font-size:34px;line-height:1.15;letter-spacing:0}.hero p{max-width:840px;margin:16px 0 0;color:#e7efff;font-size:15px}.metrics{display:grid;grid-template-columns:repeat(4,minmax(0,1fr));gap:12px;margin-top:22px}.metric{padding:14px 16px;border-radius:12px;background:rgba(255,255,255,.12);border:1px solid rgba(255,255,255,.18)}.metric span{display:block;color:#cfe1ff;font-size:12px}.metric b{display:block;font-size:22px;margin-top:2px}.section{margin-top:18px;padding:26px 28px}h2{margin:0 0 14px;font-size:22px}h3{margin:0;font-size:16px}.callout{padding:15px 16px;border-radius:12px;border:1px solid var(--line);background:#f8fbff}.callout.strong{border-color:#c9dcff;background:var(--blue-soft)}ul{margin:10px 0 0;padding-left:20px}li{margin:6px 0}.group-lists{display:grid;grid-template-columns:1fr 1fr;gap:14px}.compare-grid{display:grid;grid-template-columns:repeat(2,minmax(0,1fr));gap:14px}.compare{border:1px solid var(--line);border-radius:14px;padding:16px;background:#fff}.compare.ok{border-color:#bde5d0}.compare.warn{border-color:#ffd5a8}.compare-head{display:flex;justify-content:space-between;gap:12px;align-items:start}.chip{display:inline-flex;margin:0 0 4px 5px;padding:3px 8px;border-radius:999px;font-size:12px;border:1px solid var(--line);color:#46536a}.chip.ok{color:var(--green);background:var(--green-soft);border-color:#bde5d0}.chip.warn{color:var(--orange);background:var(--orange-soft);border-color:#ffd5a8}.delta-grid{display:grid;grid-template-columns:repeat(4,minmax(0,1fr));gap:8px;margin:14px 0}.delta-grid div{padding:10px;border-radius:10px;background:#f7f9fc}.delta-grid span{display:block;color:var(--muted);font-size:11px}.delta-grid b{font-size:18px}.good{color:var(--green)}.bad{color:var(--red)}.flat{color:#5f6c80}table{width:100%;border-collapse:collapse}th,td{padding:9px 10px;border-bottom:1px solid var(--line);text-align:right;white-space:nowrap}th:first-child,td:first-child{text-align:left}th{color:#536074;font-size:12px;background:#f7f9fc}.mini th,.mini td{padding:7px 8px;font-size:12px}.table-wrap{overflow:auto;border:1px solid var(--line);border-radius:12px}@media(max-width:980px){.shell{display:block}.side{position:static;margin-bottom:16px}.metrics,.compare-grid,.group-lists,.delta-grid{grid-template-columns:1fr}h1{font-size:28px}}
+</style>
+</head>
+<body>
+<div class="shell">
+  <aside class="side">
+    <div class="brand"><div class="brand-mark">FR</div><div><strong>FishROV Benchmark</strong><span>Android_2 复测</span></div></div>
+    <nav><a href="#summary">结论</a><a href="#quality">样本质量</a><a href="#compare">控制变量对比</a><a href="#raw">完整数据</a><a href="#next">下一轮测试</a></nav>
+  </aside>
+  <main>
+    <section class="hero"><p class="eyebrow">2026-06-15 / 百元机 / XZ2D</p><h1>A Pro VelocityLike 复测报告</h1><p>本页分析 <code>C:/Users/13999/Desktop/logs/android_2</code> 下 27 个 CSV。统计时排除了 <code>timeSeconds=0</code> 的启动样本，避免初始化尖峰污染稳态数据。</p><div class="metrics"><div class="metric"><span>CSV</span><b>27</b></div><div class="metric"><span>场景</span><b>3</b></div><div class="metric"><span>鱼数范围</span><b>300-1000</b></div><div class="metric"><span>屏幕</span><b>720x1600</b></div></div></section>
+    <section class="section" id="summary"><h2>结论</h2><div class="callout strong"><ul><li>300 鱼鱼饵争抢是最干净的样本：VelocityLike 平均 FPS 29.9，比 RVO2 高 +3.2%，比旧 A Pro 高 +2.1%；最小间距从旧 A Pro -0.625 改到 -0.141，重叠深度明显收敛。</li><li>300 鱼圆阵高压中，A Pro 两种接法 FPS 接近满帧，但平均重叠约 337.5/340.0，仍高于 RVO2 的 159.0；这更像算法/场景拥挤上限，不只是远目标语义问题。</li><li>500 鱼鱼饵争抢三组速度一致但速度倍率已到 maxSpeed 16.8：VelocityLike FPS 27.7，略好于 RVO2/旧 A Pro；重叠与旧 A Pro 基本同级，说明高速度争抢下输入语义已不是主因。</li><li>CrossFlow 500 不能直接判 VelocityLike：VelocityLike 的 maxSpeed 是 9.6，RVO2/旧 A Pro 是 2.4/2.4，速度变量被改变。</li></ul></div></section>
+    <section class="section" id="quality"><h2>样本质量</h2><div class="group-lists"><div class="callout"><strong>可以优先采信</strong><ul><li>鱼饵争抢 / 300 鱼</li><li>鱼饵争抢 / 500 鱼</li><li>鱼饵争抢 / 800 鱼</li><li>圆阵高压 / 1000 鱼</li></ul></div><div class="callout"><strong>需要谨慎</strong><ul><li>鱼饵争抢 / 1000 鱼：三后端齐全 不满足</li><li>对向穿流 / 300 鱼：三后端齐全 不满足</li><li>对向穿流 / 500 鱼：速度一致 不满足</li><li>圆阵高压 / 300 鱼：时长充足 不满足</li><li>圆阵高压 / 500 鱼：时长充足 不满足</li><li>圆阵高压 / 800 鱼：时长充足 不满足</li></ul></div></div></section>
+    <section class="section" id="compare"><h2>控制变量对比</h2><div class="compare-grid"><article class="compare warn">
+    <div class="compare-head"><h3>鱼饵争抢 / 1000 鱼</h3><div><span class="chip warn">后端不齐</span><span class="chip ok">速度一致</span><span class="chip ok">时长充足</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="">—</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="">—</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="flat">+2.2%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="good">-3.8%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>A Pro RVO</td><td>13.4</td><td>53.3</td><td>1240.8</td><td>-0.699</td><td>16.8</td><td>13.6s</td></tr><tr><td>A Pro VelocityLike</td><td>13.7</td><td>55.0</td><td>1193.7</td><td>-0.697</td><td>16.8</td><td>20.5s</td></tr></tbody></table>
+  </article>
+<article class="compare ok">
+    <div class="compare-head"><h3>鱼饵争抢 / 300 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip ok">速度一致</span><span class="chip ok">时长充足</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="good">+3.2%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+9.9%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="flat">+2.1%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="good">-3.2%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>29.0</td><td>16.8</td><td>125.3</td><td>-0.624</td><td>4.2</td><td>25.0s</td></tr><tr><td>A Pro RVO</td><td>29.3</td><td>13.9</td><td>142.1</td><td>-0.625</td><td>4.2</td><td>20.5s</td></tr><tr><td>A Pro VelocityLike</td><td>29.9</td><td>13.2</td><td>137.7</td><td>-0.141</td><td>4.2</td><td>21.0s</td></tr></tbody></table>
+  </article>
+<article class="compare ok">
+    <div class="compare-head"><h3>鱼饵争抢 / 500 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip ok">速度一致</span><span class="chip ok">时长充足</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="good">+3.2%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+93.8%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="flat">+2.9%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="flat">-1.0%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>26.8</td><td>27.3</td><td>277.1</td><td>-0.682</td><td>16.8</td><td>8.0s</td></tr><tr><td>A Pro RVO</td><td>26.9</td><td>24.6</td><td>542.2</td><td>-0.684</td><td>16.8</td><td>13.5s</td></tr><tr><td>A Pro VelocityLike</td><td>27.7</td><td>23.6</td><td>536.9</td><td>-0.678</td><td>16.8</td><td>27.5s</td></tr></tbody></table>
+  </article>
+<article class="compare ok">
+    <div class="compare-head"><h3>鱼饵争抢 / 800 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip ok">速度一致</span><span class="chip ok">时长充足</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="good">+22.8%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+67.6%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="good">+11.6%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="bad">+3.2%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>16.8</td><td>47.5</td><td>640.1</td><td>-0.694</td><td>16.8</td><td>10.5s</td></tr><tr><td>A Pro RVO</td><td>18.5</td><td>39.9</td><td>1040.4</td><td>-0.697</td><td>16.8</td><td>15.0s</td></tr><tr><td>A Pro VelocityLike</td><td>20.6</td><td>39.9</td><td>1073.2</td><td>-0.694</td><td>16.8</td><td>17.5s</td></tr></tbody></table>
+  </article>
+<article class="compare warn">
+    <div class="compare-head"><h3>对向穿流 / 300 鱼</h3><div><span class="chip warn">后端不齐</span><span class="chip ok">速度一致</span><span class="chip ok">时长充足</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="">—</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="">—</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="">—</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="">—</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>30.0</td><td>16.1</td><td>82.2</td><td>-0.268</td><td>2.4</td><td>9.5s</td></tr></tbody></table>
+  </article>
+<article class="compare warn">
+    <div class="compare-head"><h3>对向穿流 / 500 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip warn">速度不一致</span><span class="chip ok">时长充足</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="good">+7.1%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+58.4%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="flat">+0.8%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="bad">+22.9%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>26.4</td><td>26.8</td><td>171.2</td><td>-0.398</td><td>2.4</td><td>20.5s</td></tr><tr><td>A Pro RVO</td><td>28.1</td><td>22.1</td><td>220.8</td><td>-0.459</td><td>2.4</td><td>28.0s</td></tr><tr><td>A Pro VelocityLike</td><td>28.3</td><td>23.4</td><td>271.3</td><td>-0.618</td><td>9.6</td><td>25.0s</td></tr></tbody></table>
+  </article>
+<article class="compare ok">
+    <div class="compare-head"><h3>圆阵高压 / 1000 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip ok">速度一致</span><span class="chip ok">时长充足</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="good">+5.5%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+30.1%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="flat">+1.6%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="flat">-0.4%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>13.4</td><td>56.7</td><td>1002.1</td><td>-0.688</td><td>9.6</td><td>9.5s</td></tr><tr><td>A Pro RVO</td><td>13.9</td><td>56.1</td><td>1308.8</td><td>-0.697</td><td>9.6</td><td>25.1s</td></tr><tr><td>A Pro VelocityLike</td><td>14.1</td><td>53.4</td><td>1304.0</td><td>-0.692</td><td>9.6</td><td>22.5s</td></tr></tbody></table>
+  </article>
+<article class="compare warn">
+    <div class="compare-head"><h3>圆阵高压 / 300 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip ok">速度一致</span><span class="chip warn">时长偏短</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="flat">-0.7%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+113.8%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="flat">-0.5%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="flat">+0.7%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>30.1</td><td>16.7</td><td>159.0</td><td>-0.379</td><td>2.4</td><td>14.5s</td></tr><tr><td>A Pro RVO</td><td>30.0</td><td>12.5</td><td>337.5</td><td>-0.383</td><td>2.4</td><td>8.0s</td></tr><tr><td>A Pro VelocityLike</td><td>29.9</td><td>12.9</td><td>340.0</td><td>-0.382</td><td>2.4</td><td>7.5s</td></tr></tbody></table>
+  </article>
+<article class="compare warn">
+    <div class="compare-head"><h3>圆阵高压 / 500 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip ok">速度一致</span><span class="chip warn">时长偏短</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="good">+3.6%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+4.2%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="flat">+1.4%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="good">-16.7%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>28.3</td><td>22.3</td><td>600.0</td><td>-0.416</td><td>9.6</td><td>2.5s</td></tr><tr><td>A Pro RVO</td><td>28.9</td><td>19.1</td><td>750.0</td><td>-0.416</td><td>9.6</td><td>1.0s</td></tr><tr><td>A Pro VelocityLike</td><td>29.3</td><td>22.8</td><td>625.0</td><td>-0.416</td><td>9.6</td><td>2.0s</td></tr></tbody></table>
+  </article>
+<article class="compare warn">
+    <div class="compare-head"><h3>圆阵高压 / 800 鱼</h3><div><span class="chip ok">三后端齐全</span><span class="chip ok">速度一致</span><span class="chip warn">时长偏短</span></div></div>
+    <div class="delta-grid">
+      <div><span>Velocity vs RVO2 FPS</span><b class="good">+17.6%</b></div>
+      <div><span>Velocity vs RVO2 重叠</span><b class="bad">+327.0%</b></div>
+      <div><span>Velocity vs 旧 A Pro FPS</span><b class="bad">-4.3%</b></div>
+      <div><span>Velocity vs 旧 A Pro 重叠</span><b class="flat">+0.0%</b></div>
+    </div>
+    <table class="mini"><thead><tr><th>后端</th><th>FPS</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>MaxSpeed</th><th>时长</th></tr></thead><tbody><tr><td>RVO2</td><td>15.1</td><td>77.9</td><td>562.1</td><td>-0.697</td><td>9.6</td><td>9.5s</td></tr><tr><td>A Pro RVO</td><td>18.6</td><td>39.8</td><td>2400.0</td><td>-0.511</td><td>9.6</td><td>2.5s</td></tr><tr><td>A Pro VelocityLike</td><td>17.8</td><td>35.8</td><td>2400.0</td><td>-0.521</td><td>9.6</td><td>4.5s</td></tr></tbody></table>
+  </article></div></section>
+    <section class="section" id="raw"><h2>完整汇总表</h2><div class="table-wrap"><table><thead><tr><th>场景</th><th>鱼数</th><th>后端</th><th>MaxSpeed</th><th>样本</th><th>时长</th><th>FPS</th><th>1% Low</th><th>Sim ms</th><th>重叠</th><th>最小间距</th><th>均速</th><th>均加速</th></tr></thead><tbody><tr><td>鱼饵争抢</td><td>300</td><td>RVO2</td><td>4.2</td><td>50</td><td>25.0s</td><td>29.0</td><td>17.6</td><td>16.8</td><td>125.3</td><td>-0.624</td><td>1.9</td><td>4.0</td></tr>
+<tr><td>鱼饵争抢</td><td>300</td><td>A Pro RVO</td><td>4.2</td><td>41</td><td>20.5s</td><td>29.3</td><td>13.9</td><td>13.9</td><td>142.1</td><td>-0.625</td><td>1.8</td><td>3.9</td></tr>
+<tr><td>鱼饵争抢</td><td>300</td><td>A Pro VelocityLike</td><td>4.2</td><td>42</td><td>21.0s</td><td>29.9</td><td>20.5</td><td>13.2</td><td>137.7</td><td>-0.141</td><td>1.8</td><td>3.4</td></tr>
+<tr><td>鱼饵争抢</td><td>500</td><td>RVO2</td><td>16.8</td><td>16</td><td>8.0s</td><td>26.8</td><td>8.5</td><td>27.3</td><td>277.1</td><td>-0.682</td><td>10.5</td><td>54.1</td></tr>
+<tr><td>鱼饵争抢</td><td>500</td><td>A Pro RVO</td><td>16.8</td><td>27</td><td>13.5s</td><td>26.9</td><td>11.3</td><td>24.6</td><td>542.2</td><td>-0.684</td><td>5.3</td><td>23.0</td></tr>
+<tr><td>鱼饵争抢</td><td>500</td><td>A Pro VelocityLike</td><td>16.8</td><td>55</td><td>27.5s</td><td>27.7</td><td>13.7</td><td>23.6</td><td>536.9</td><td>-0.678</td><td>5.1</td><td>21.2</td></tr>
+<tr><td>鱼饵争抢</td><td>800</td><td>RVO2</td><td>16.8</td><td>21</td><td>10.5s</td><td>16.8</td><td>6.1</td><td>47.5</td><td>640.1</td><td>-0.694</td><td>9.4</td><td>35.3</td></tr>
+<tr><td>鱼饵争抢</td><td>800</td><td>A Pro RVO</td><td>16.8</td><td>30</td><td>15.0s</td><td>18.5</td><td>5.8</td><td>39.9</td><td>1040.4</td><td>-0.697</td><td>5.6</td><td>26.7</td></tr>
+<tr><td>鱼饵争抢</td><td>800</td><td>A Pro VelocityLike</td><td>16.8</td><td>35</td><td>17.5s</td><td>20.6</td><td>6.5</td><td>39.9</td><td>1073.2</td><td>-0.694</td><td>5.4</td><td>24.6</td></tr>
+<tr><td>鱼饵争抢</td><td>1000</td><td>A Pro RVO</td><td>16.8</td><td>27</td><td>13.6s</td><td>13.4</td><td>4.7</td><td>53.3</td><td>1240.8</td><td>-0.699</td><td>6.0</td><td>26.9</td></tr>
+<tr><td>鱼饵争抢</td><td>1000</td><td>A Pro VelocityLike</td><td>16.8</td><td>41</td><td>20.5s</td><td>13.7</td><td>5.4</td><td>55.0</td><td>1193.7</td><td>-0.697</td><td>6.0</td><td>26.9</td></tr>
+<tr><td>对向穿流</td><td>300</td><td>RVO2</td><td>2.4</td><td>19</td><td>9.5s</td><td>30.0</td><td>21.2</td><td>16.1</td><td>82.2</td><td>-0.268</td><td>1.8</td><td>1.9</td></tr>
+<tr><td>对向穿流</td><td>500</td><td>RVO2</td><td>2.4</td><td>41</td><td>20.5s</td><td>26.4</td><td>11.8</td><td>26.8</td><td>171.2</td><td>-0.398</td><td>1.5</td><td>1.3</td></tr>
+<tr><td>对向穿流</td><td>500</td><td>A Pro RVO</td><td>2.4</td><td>56</td><td>28.0s</td><td>28.1</td><td>14.2</td><td>22.1</td><td>220.8</td><td>-0.459</td><td>1.4</td><td>1.1</td></tr>
+<tr><td>对向穿流</td><td>500</td><td>A Pro VelocityLike</td><td>9.6</td><td>50</td><td>25.0s</td><td>28.3</td><td>14.1</td><td>23.4</td><td>271.3</td><td>-0.618</td><td>2.4</td><td>5.7</td></tr>
+<tr><td>圆阵高压</td><td>300</td><td>RVO2</td><td>2.4</td><td>29</td><td>14.5s</td><td>30.1</td><td>24.1</td><td>16.7</td><td>159.0</td><td>-0.379</td><td>1.5</td><td>1.7</td></tr>
+<tr><td>圆阵高压</td><td>300</td><td>A Pro RVO</td><td>2.4</td><td>16</td><td>8.0s</td><td>30.0</td><td>20.7</td><td>12.5</td><td>337.5</td><td>-0.383</td><td>1.6</td><td>0.0</td></tr>
+<tr><td>圆阵高压</td><td>300</td><td>A Pro VelocityLike</td><td>2.4</td><td>15</td><td>7.5s</td><td>29.9</td><td>16.9</td><td>12.9</td><td>340.0</td><td>-0.382</td><td>1.8</td><td>0.0</td></tr>
+<tr><td>圆阵高压</td><td>500</td><td>RVO2</td><td>9.6</td><td>5</td><td>2.5s</td><td>28.3</td><td>7.5</td><td>22.3</td><td>600.0</td><td>-0.416</td><td>9.6</td><td>0.0</td></tr>
+<tr><td>圆阵高压</td><td>500</td><td>A Pro RVO</td><td>9.6</td><td>2</td><td>1.0s</td><td>28.9</td><td>9.5</td><td>19.1</td><td>750.0</td><td>-0.416</td><td>4.8</td><td>0.0</td></tr>
+<tr><td>圆阵高压</td><td>500</td><td>A Pro VelocityLike</td><td>9.6</td><td>4</td><td>2.0s</td><td>29.3</td><td>12.9</td><td>22.8</td><td>625.0</td><td>-0.416</td><td>2.4</td><td>0.0</td></tr>
+<tr><td>圆阵高压</td><td>800</td><td>RVO2</td><td>9.6</td><td>19</td><td>9.5s</td><td>15.1</td><td>4.4</td><td>77.9</td><td>562.1</td><td>-0.697</td><td>6.9</td><td>13.8</td></tr>
+<tr><td>圆阵高压</td><td>800</td><td>A Pro RVO</td><td>9.6</td><td>5</td><td>2.5s</td><td>18.6</td><td>5.8</td><td>39.8</td><td>2400.0</td><td>-0.511</td><td>1.9</td><td>0.1</td></tr>
+<tr><td>圆阵高压</td><td>800</td><td>A Pro VelocityLike</td><td>9.6</td><td>9</td><td>4.5s</td><td>17.8</td><td>6.7</td><td>35.8</td><td>2400.0</td><td>-0.521</td><td>1.1</td><td>0.0</td></tr>
+<tr><td>圆阵高压</td><td>1000</td><td>RVO2</td><td>9.6</td><td>19</td><td>9.5s</td><td>13.4</td><td>3.2</td><td>56.7</td><td>1002.1</td><td>-0.688</td><td>7.2</td><td>11.5</td></tr>
+<tr><td>圆阵高压</td><td>1000</td><td>A Pro RVO</td><td>9.6</td><td>50</td><td>25.1s</td><td>13.9</td><td>5.5</td><td>56.1</td><td>1308.8</td><td>-0.697</td><td>2.8</td><td>7.6</td></tr>
+<tr><td>圆阵高压</td><td>1000</td><td>A Pro VelocityLike</td><td>9.6</td><td>45</td><td>22.5s</td><td>14.1</td><td>4.6</td><td>53.4</td><td>1304.0</td><td>-0.692</td><td>2.9</td><td>8.2</td></tr></tbody></table></div></section>
+    <section class="section" id="next"><h2>下一轮测试建议</h2><div class="callout"><ul><li>固定速度倍率为 1x 后重新跑 CrossFlow 500，因为这一组 VelocityLike 的 MaxSpeed 与另外两个后端不一致。</li><li>Dense 500 / Dense 800 每个后端至少跑到 20 秒；当前几组只有 1-5 秒，重叠和 1% Low 都容易被启动阶段支配。</li><li>BaitScramble 建议保留 300、500、800、1000 四档，但同一档必须连续跑完三后端，不要中途改速度倍率。</li><li>如果目标是判断 A Pro 能不能用，首要看 VelocityLike 在 300/500 鱼下是否同时满足 FPS 更高、最小间距不恶化、重叠不显著高于 RVO2。</li></ul></div></section>
+  </main>
+</div>
+</body>
+</html>

FishROV/Docs/avoidance-benchmark-report-20260615.html

@@ -0,0 +1,562 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <meta name="color-scheme" content="light">
+  <title>FishROV A Pro RVO vs RVO2 Benchmark 报告</title>
+  <style>
+    :root {
+      --ink: #172033;
+      --muted: #667085;
+      --paper: #ffffff;
+      --canvas: #eef2f8;
+      --line: #dce3ee;
+      --navy: #102a43;
+      --blue: #2f6fed;
+      --blue-soft: #eaf1ff;
+      --cyan: #13a8a8;
+      --cyan-soft: #e6fafa;
+      --green: #138a5b;
+      --green-soft: #e7f7ef;
+      --orange: #e77817;
+      --orange-soft: #fff2e5;
+      --red: #c73e4d;
+      --red-soft: #ffeaed;
+      --shadow: 0 14px 40px rgba(16, 42, 67, .09);
+    }
+    * { box-sizing: border-box; }
+    html { scroll-behavior: smooth; }
+    body {
+      margin: 0;
+      color: var(--ink);
+      background:
+        radial-gradient(circle at 8% 0%, rgba(47, 111, 237, .10), transparent 28rem),
+        linear-gradient(180deg, #f7f9fc 0, var(--canvas) 100%);
+      font: 14px/1.7 -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC", "Microsoft YaHei", sans-serif;
+    }
+    a { color: inherit; }
+    .shell {
+      display: grid;
+      grid-template-columns: 248px minmax(0, 1180px);
+      gap: 26px;
+      width: min(1480px, calc(100% - 32px));
+      margin: 24px auto 64px;
+      align-items: start;
+    }
+    .side {
+      position: sticky;
+      top: 20px;
+      min-width: 0;
+      padding: 20px 16px;
+      border: 1px solid var(--line);
+      border-radius: 16px;
+      background: rgba(255, 255, 255, .94);
+      box-shadow: var(--shadow);
+      backdrop-filter: blur(10px);
+    }
+    .brand {
+      display: flex;
+      gap: 11px;
+      align-items: center;
+      padding: 2px 5px 17px;
+      border-bottom: 1px solid var(--line);
+    }
+    .brand-mark {
+      display: grid;
+      width: 38px;
+      height: 38px;
+      place-items: center;
+      border-radius: 11px;
+      color: #fff;
+      background: linear-gradient(145deg, var(--blue), #1744ad);
+      font-size: 18px;
+      font-weight: 800;
+      box-shadow: 0 8px 16px rgba(47, 111, 237, .28);
+    }
+    .brand strong { display: block; font-size: 15px; }
+    .brand span { display: block; color: var(--muted); font-size: 11px; }
+    .side nav { margin-top: 14px; }
+    .side nav a {
+      display: block;
+      padding: 7px 10px;
+      border-radius: 8px;
+      color: #46536a;
+      text-decoration: none;
+      font-size: 13px;
+    }
+    .side nav a:hover { color: var(--blue); background: var(--blue-soft); }
+    .side-note {
+      margin: 15px 5px 0;
+      padding-top: 14px;
+      border-top: 1px solid var(--line);
+      color: var(--muted);
+      font-size: 11px;
+    }
+    main { min-width: 0; }
+    .hero, .section {
+      border: 1px solid var(--line);
+      border-radius: 18px;
+      background: var(--paper);
+      box-shadow: var(--shadow);
+    }
+    .hero {
+      position: relative;
+      overflow: hidden;
+      padding: 38px 40px 34px;
+      color: #fff;
+      background:
+        linear-gradient(115deg, rgba(16, 42, 67, .98), rgba(23, 64, 112, .96) 56%, rgba(47, 111, 237, .92)),
+        #102a43;
+    }
+    .hero::after {
+      position: absolute;
+      right: -55px;
+      bottom: -115px;
+      width: 330px;
+      height: 330px;
+      border: 45px solid rgba(255, 255, 255, .055);
+      border-radius: 50%;
+      content: "";
+    }
+    .eyebrow {
+      margin: 0 0 10px;
+      color: #a8c8ff;
+      font-size: 12px;
+      font-weight: 750;
+      letter-spacing: .13em;
+      text-transform: uppercase;
+    }
+    h1 {
+      max-width: 860px;
+      margin: 0;
+      font-size: clamp(30px, 5vw, 48px);
+      line-height: 1.18;
+      letter-spacing: -.025em;
+    }
+    .lead {
+      max-width: 860px;
+      margin: 16px 0 0;
+      color: #d8e6fb;
+      font-size: 16px;
+    }
+    .hero-meta, .summary-grid, .grid-2, .grid-3 {
+      display: grid;
+      gap: 14px;
+    }
+    .hero-meta {
+      grid-template-columns: repeat(4, minmax(0, 1fr));
+      margin-top: 28px;
+    }
+    .hero-meta div {
+      padding: 12px 14px;
+      border: 1px solid rgba(255, 255, 255, .13);
+      border-radius: 11px;
+      background: rgba(255, 255, 255, .07);
+    }
+    .hero-meta span { display: block; color: #a9bdd8; font-size: 11px; }
+    .hero-meta strong { display: block; margin-top: 2px; font-size: 14px; }
+    .section {
+      margin-top: 20px;
+      padding: 28px 30px;
+      scroll-margin-top: 24px;
+    }
+    .section-head {
+      display: flex;
+      gap: 20px;
+      align-items: flex-start;
+      justify-content: space-between;
+      margin-bottom: 20px;
+    }
+    .kicker {
+      margin: 0 0 3px;
+      color: var(--blue);
+      font-size: 11px;
+      font-weight: 800;
+      letter-spacing: .12em;
+      text-transform: uppercase;
+    }
+    h2 { margin: 0; color: var(--navy); font-size: 24px; line-height: 1.3; }
+    h3 { margin: 0 0 10px; color: var(--navy); font-size: 16px; }
+    p { margin: 8px 0; }
+    .muted { color: var(--muted); }
+    .summary-grid { grid-template-columns: repeat(4, minmax(0, 1fr)); }
+    .metric {
+      padding: 17px;
+      border: 1px solid var(--line);
+      border-radius: 13px;
+      background: #fbfcfe;
+    }
+    .metric b { display: block; color: var(--blue); font-size: 27px; line-height: 1.2; }
+    .metric span { display: block; margin-top: 6px; color: var(--muted); font-size: 12px; }
+    .grid-2 { grid-template-columns: repeat(2, minmax(0, 1fr)); }
+    .grid-3 { grid-template-columns: repeat(3, minmax(0, 1fr)); }
+    .grid-2 > *, .grid-3 > * { min-width: 0; }
+    .info-card, .compare-card {
+      padding: 17px 18px;
+      border: 1px solid var(--line);
+      border-radius: 13px;
+      background: #fbfcfe;
+    }
+    .compare-card .rule {
+      margin-bottom: 13px;
+      padding-bottom: 11px;
+      border-bottom: 1px solid var(--line);
+      color: var(--muted);
+      font-size: 12px;
+    }
+    .compare-card .rule strong { color: var(--navy); }
+    .tag {
+      display: inline-block;
+      margin: 2px 5px 2px 0;
+      padding: 3px 9px;
+      border-radius: 999px;
+      color: #31507d;
+      background: var(--blue-soft);
+      font-size: 11px;
+      font-weight: 700;
+    }
+    .tag.green { color: var(--green); background: var(--green-soft); }
+    .tag.orange { color: #9a4e05; background: var(--orange-soft); }
+    .tag.red { color: var(--red); background: var(--red-soft); }
+    .tag.cyan { color: #087273; background: var(--cyan-soft); }
+    .callout {
+      margin: 17px 0 0;
+      padding: 14px 16px;
+      border-left: 4px solid var(--blue);
+      border-radius: 9px;
+      background: var(--blue-soft);
+    }
+    .callout.warn { border-color: var(--orange); background: var(--orange-soft); }
+    .callout.bad { border-color: var(--red); background: var(--red-soft); }
+    .callout.good { border-color: var(--green); background: var(--green-soft); }
+    table {
+      width: 100%;
+      border-collapse: collapse;
+      font-size: 13px;
+    }
+    th, td {
+      padding: 10px 9px;
+      border-bottom: 1px solid var(--line);
+      text-align: right;
+      vertical-align: top;
+      white-space: nowrap;
+      font-variant-numeric: tabular-nums;
+    }
+    th {
+      color: #31507d;
+      background: #f3f7fd;
+      font-size: 12px;
+      font-weight: 800;
+    }
+    th:first-child, td:first-child,
+    th:nth-child(2), td:nth-child(2) { text-align: left; }
+    .wide-table { overflow-x: auto; }
+    .score {
+      display: inline-flex;
+      min-width: 68px;
+      justify-content: center;
+      padding: 2px 8px;
+      border-radius: 999px;
+      font-size: 12px;
+      font-weight: 800;
+    }
+    .score.good { color: var(--green); background: var(--green-soft); }
+    .score.warn { color: #9a4e05; background: var(--orange-soft); }
+    .score.bad { color: var(--red); background: var(--red-soft); }
+    .good-text { color: var(--green); font-weight: 800; }
+    .warn-text { color: var(--orange); font-weight: 800; }
+    .bad-text { color: var(--red); font-weight: 800; }
+    .bar-list {
+      display: grid;
+      gap: 10px;
+      margin-top: 12px;
+    }
+    .bar-row {
+      display: grid;
+      grid-template-columns: 110px 1fr 78px;
+      gap: 10px;
+      align-items: center;
+      font-size: 12px;
+      font-weight: 750;
+      font-variant-numeric: tabular-nums;
+    }
+    .track {
+      height: 14px;
+      border-radius: 999px;
+      background: #e8eef7;
+      overflow: hidden;
+    }
+    .fill {
+      height: 100%;
+      border-radius: 999px;
+      background: var(--blue);
+    }
+    .fill.green { background: var(--green); }
+    .fill.orange { background: var(--orange); }
+    .fill.red { background: var(--red); }
+    .foot {
+      margin-top: 18px;
+      color: var(--muted);
+      font-size: 12px;
+    }
+    @media (max-width: 1080px) {
+      .shell { grid-template-columns: 1fr; }
+      .side { position: static; }
+      .side nav { display: grid; grid-template-columns: repeat(2, minmax(0, 1fr)); }
+      .hero-meta, .summary-grid, .grid-2, .grid-3 { grid-template-columns: 1fr; }
+      .section-head { display: block; }
+    }
+    @media (max-width: 640px) {
+      .shell { width: min(100% - 20px, 1480px); margin-top: 10px; }
+      .hero, .section { border-radius: 14px; }
+      .hero { padding: 28px 22px; }
+      .section { padding: 22px 18px; }
+      .side nav { grid-template-columns: 1fr; }
+      .bar-row { grid-template-columns: 90px 1fr 64px; }
+    }
+  </style>
+</head>
+<body>
+  <div class="shell">
+    <aside class="side">
+      <div class="brand">
+        <div class="brand-mark">AP</div>
+        <div>
+          <strong>A Pro RVO 报告</strong>
+          <span>固定 XZ2D / 鱼饵场景纳入</span>
+        </div>
+      </div>
+      <nav>
+        <a href="#summary">结论摘要</a>
+        <a href="#android">百元机：A Pro vs RVO2</a>
+        <a href="#bait">鱼饵场景</a>
+        <a href="#editor">编辑器趋势</a>
+        <a href="#env">跨环境差异</a>
+        <a href="#method">样本与口径</a>
+        <a href="#actions">建议动作</a>
+      </nav>
+      <div class="side-note">
+        数据源：C:\Users\13999\Desktop\logs。统计稳定段，排除 timeSeconds = 0 的启动采样。
+      </div>
+    </aside>
+
+    <main>
+      <section class="hero">
+        <p class="eyebrow">2026-06-15 / Android + Editor / XZ2D mainline</p>
+        <h1>A Pro RVO 提升性能，但在 2D 下常用重叠换速度</h1>
+        <p class="lead">这批新日志固定了大部分 2D 主线，新增了 BaitScramble 鱼饵场景。结论不是“A Pro 全面替代 RVO2”：它在百元机 500 鱼附近明显改善 FPS 和 simulation ms，但鱼饵、FreeSwim、Dense 的空间重叠通常更高。</p>
+        <div class="hero-meta">
+          <div><span>日志数量</span><strong>69 个 CSV</strong></div>
+          <div><span>主环境</span><strong>Android 百元机</strong></div>
+          <div><span>主维度</span><strong>XZ2D</strong></div>
+          <div><span>新增场景</span><strong>BaitScramble</strong></div>
+        </div>
+      </section>
+
+      <section id="summary" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Executive Summary</p>
+            <h2>结论摘要</h2>
+          </div>
+          <p class="muted">这次应该把“帧率胜利”和“避障质量胜利”分开看。</p>
+        </div>
+        <div class="summary-grid">
+          <div class="metric"><b>500</b><span>A Pro 的价值最明显：百元机 500 鱼在 FreeSwim / CrossFlow / Dense / Bait 都显著提升 FPS。</span></div>
+          <div class="metric"><b>+306</b><span>鱼饵 500 鱼 A Pro 比 RVO2 多 306.3 个平均空间重叠对，性能提升伴随拥挤代价。</span></div>
+          <div class="metric"><b>Cross 500</b><span>百元机 CrossFlow 500 是 A Pro 最像“真胜利”的样本：FPS 28.8 vs 11.4，空间重叠 288.7 vs 512.0。</span></div>
+          <div class="metric"><b>800+</b><span>A Pro 到 800/1000 后重叠经常爆高，不能直接作为高数量默认方案。</span></div>
+        </div>
+        <div class="callout warn">
+          <strong>推荐判断：</strong>A Pro 可以继续作为性能候选，但现在还不建议无条件替代 RVO2。鱼饵场景要先解决聚集重叠或增加后处理分离，再谈默认启用。
+        </div>
+      </section>
+
+      <section id="android" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Android / Same Config</p>
+            <h2>百元机：A Pro vs RVO2</h2>
+          </div>
+          <p class="muted">固定环境、场景、维度和数量，只比较模式。负数 Sim Δ 表示 A Pro 更快；正数重叠 Δ 表示 A Pro 更拥挤。</p>
+        </div>
+
+        <div class="wide-table compare-card">
+          <div class="rule"><strong>固定：</strong>Android / XZ2D / 同场景 / 同数量。<strong>唯一变量：</strong>RVO2 vs A Pro RVO。</div>
+          <table>
+            <thead>
+              <tr>
+                <th>场景</th>
+                <th>数量</th>
+                <th>FPS RVO2</th>
+                <th>FPS A Pro</th>
+                <th>FPS Δ</th>
+                <th>Sim Δ ms</th>
+                <th>空间重叠 RVO2</th>
+                <th>空间重叠 A Pro</th>
+                <th>重叠 Δ</th>
+                <th>结论</th>
+              </tr>
+            </thead>
+            <tbody>
+              <tr><td>FreeSwim</td><td>100</td><td>22.0</td><td>30.2</td><td class="good-text">+8.2</td><td class="warn-text">+0.26</td><td>0.4</td><td>0.4</td><td>0.0</td><td><span class="score good">可用</span></td></tr>
+              <tr><td>FreeSwim</td><td>300</td><td>22.5</td><td>30.0</td><td class="good-text">+7.5</td><td class="good-text">-7.08</td><td>16.1</td><td>17.9</td><td class="warn-text">+1.8</td><td><span class="score good">小赢</span></td></tr>
+              <tr><td>FreeSwim</td><td>500</td><td>11.9</td><td>28.3</td><td class="good-text">+16.4</td><td class="good-text">-22.89</td><td>89.1</td><td>268.0</td><td class="bad-text">+178.9</td><td><span class="score warn">性能赢</span></td></tr>
+              <tr><td>FreeSwim</td><td>800</td><td>19.0</td><td>21.7</td><td class="good-text">+2.7</td><td class="good-text">-17.21</td><td>293.1</td><td>844.3</td><td class="bad-text">+551.2</td><td><span class="score bad">质量差</span></td></tr>
+              <tr><td>CrossFlow</td><td>500</td><td>11.4</td><td>28.8</td><td class="good-text">+17.4</td><td class="good-text">-50.76</td><td>512.0</td><td>288.7</td><td class="good-text">-223.3</td><td><span class="score good">真胜利</span></td></tr>
+              <tr><td>Dense</td><td>500</td><td>13.0</td><td>27.9</td><td class="good-text">+14.9</td><td class="good-text">-37.78</td><td>625.0</td><td>545.5</td><td class="good-text">-79.5</td><td><span class="score good">可试</span></td></tr>
+              <tr><td>Dense</td><td>800</td><td>13.9</td><td>19.1</td><td class="good-text">+5.2</td><td class="good-text">-35.02</td><td>525.7</td><td>2400.0</td><td class="bad-text">+1874.3</td><td><span class="score bad">重叠爆</span></td></tr>
+              <tr><td>BaitScramble</td><td>500</td><td>16.5</td><td>28.2</td><td class="good-text">+11.7</td><td class="good-text">-41.27</td><td>209.9</td><td>516.2</td><td class="bad-text">+306.3</td><td><span class="score warn">性能赢</span></td></tr>
+              <tr><td>BaitScramble</td><td>800</td><td>12.9</td><td>18.8</td><td class="good-text">+5.9</td><td class="good-text">-8.02</td><td>616.0</td><td>1031.9</td><td class="bad-text">+415.9</td><td><span class="score bad">质量差</span></td></tr>
+              <tr><td>BaitScramble</td><td>1000</td><td>14.0</td><td>15.5</td><td class="good-text">+1.5</td><td class="good-text">-9.80</td><td>866.9</td><td>1208.8</td><td class="bad-text">+341.9</td><td><span class="score bad">不建议</span></td></tr>
+            </tbody>
+          </table>
+        </div>
+      </section>
+
+      <section id="bait" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">BaitScramble</p>
+            <h2>鱼饵场景是关键：A Pro 帧率更好，但聚集更挤</h2>
+          </div>
+          <p class="muted">鱼饵场景更接近真实玩法，所以它的质量风险比 FreeSwim 更值得优先处理。</p>
+        </div>
+        <div class="grid-2">
+          <article class="compare-card">
+            <h3>Android / BaitScramble / XZ2D</h3>
+            <div class="rule"><strong>固定：</strong>Android / BaitScramble / XZ2D。<strong>变量：</strong>数量与模式。</div>
+            <table>
+              <thead><tr><th>数量</th><th>模式</th><th>FPS</th><th>1% Low</th><th>Sim ms</th><th>空间重叠</th><th>最小间距</th></tr></thead>
+              <tbody>
+                <tr><td>100</td><td>RVO2</td><td>30.2</td><td>29.1</td><td>7.50</td><td>12.8</td><td>-0.610</td></tr>
+                <tr><td>100</td><td>A Pro</td><td>29.1</td><td>7.5</td><td>6.15</td><td>18.2</td><td>-0.649</td></tr>
+                <tr><td>300</td><td>RVO2</td><td>26.5</td><td>8.8</td><td>33.41</td><td>122.2</td><td>-0.670</td></tr>
+                <tr><td>300</td><td>A Pro</td><td>30.1</td><td>24.8</td><td>15.62</td><td>235.9</td><td>-0.674</td></tr>
+                <tr><td>500</td><td>RVO2</td><td>16.5</td><td>3.7</td><td>65.92</td><td>209.9</td><td>-0.694</td></tr>
+                <tr><td>500</td><td>A Pro</td><td>28.2</td><td>12.8</td><td>24.65</td><td>516.2</td><td>-0.686</td></tr>
+                <tr><td>800</td><td>RVO2</td><td>12.9</td><td>3.4</td><td>52.27</td><td>616.0</td><td>-0.696</td></tr>
+                <tr><td>800</td><td>A Pro</td><td>18.8</td><td>6.0</td><td>44.25</td><td>1031.9</td><td>-0.692</td></tr>
+                <tr><td>1000</td><td>RVO2</td><td>14.0</td><td>4.6</td><td>72.77</td><td>866.9</td><td>-0.697</td></tr>
+                <tr><td>1000</td><td>A Pro</td><td>15.5</td><td>5.4</td><td>62.97</td><td>1208.8</td><td>-0.692</td></tr>
+              </tbody>
+            </table>
+          </article>
+          <article class="compare-card">
+            <h3>鱼饵场景读法</h3>
+            <div class="rule"><strong>业务解释：</strong>鱼饵会制造局部聚集，空间重叠比单纯 FPS 更关键。</div>
+            <div class="bar-list">
+              <div class="bar-row"><span>300 FPS</span><div class="track"><div class="fill green" style="width:100%"></div></div><strong>A Pro</strong></div>
+              <div class="bar-row"><span>300 重叠</span><div class="track"><div class="fill orange" style="width:52%"></div></div><strong>RVO2</strong></div>
+              <div class="bar-row"><span>500 FPS</span><div class="track"><div class="fill green" style="width:100%"></div></div><strong>A Pro</strong></div>
+              <div class="bar-row"><span>500 重叠</span><div class="track"><div class="fill red" style="width:100%"></div></div><strong>RVO2</strong></div>
+              <div class="bar-row"><span>800 FPS</span><div class="track"><div class="fill orange" style="width:67%"></div></div><strong>A Pro</strong></div>
+              <div class="bar-row"><span>800 重叠</span><div class="track"><div class="fill red" style="width:100%"></div></div><strong>RVO2</strong></div>
+            </div>
+            <div class="callout bad">
+              <strong>当前结论：</strong>A Pro 更能把鱼饵场景跑起来，但 300/500/800/1000 全部比 RVO2 更拥挤。它需要局部密度限制、到达鱼饵后的分层/环绕点、或重叠后处理。
+            </div>
+          </article>
+        </div>
+      </section>
+
+      <section id="editor" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Editor Trend</p>
+            <h2>编辑器趋势：A Pro 在大数量下也不是质量胜利</h2>
+          </div>
+          <p class="muted">编辑器只看趋势，不当移动端上线依据。</p>
+        </div>
+        <div class="wide-table compare-card">
+          <div class="rule"><strong>固定：</strong>Editor / XZ2D / 同场景 / 同数量。<strong>唯一变量：</strong>RVO2 vs A Pro。</div>
+          <table>
+            <thead><tr><th>场景</th><th>数量</th><th>FPS RVO2</th><th>FPS A Pro</th><th>Sim Δ ms</th><th>空间重叠 RVO2</th><th>空间重叠 A Pro</th><th>结论</th></tr></thead>
+            <tbody>
+              <tr><td>FreeSwim</td><td>300</td><td>99.3</td><td>100.9</td><td class="good-text">-0.22</td><td>6.0</td><td>5.8</td><td><span class="score good">基本持平</span></td></tr>
+              <tr><td>FreeSwim</td><td>3000</td><td>28.4</td><td>26.7</td><td class="good-text">-0.29</td><td>4222.1</td><td>4898.8</td><td><span class="score bad">更拥挤</span></td></tr>
+              <tr><td>BaitScramble</td><td>300</td><td>98.5</td><td>105.6</td><td class="good-text">-0.24</td><td>150.9</td><td>166.1</td><td><span class="score warn">小幅更挤</span></td></tr>
+              <tr><td>BaitScramble</td><td>3000</td><td>30.3</td><td>24.8</td><td class="bad-text">+2.25</td><td>4955.2</td><td>7677.1</td><td><span class="score bad">双输</span></td></tr>
+              <tr><td>CrossFlow</td><td>3000</td><td>28.0</td><td>26.9</td><td class="bad-text">+0.77</td><td>3906.3</td><td>6149.5</td><td><span class="score bad">更拥挤</span></td></tr>
+              <tr><td>Dense</td><td>3000</td><td>25.8</td><td>26.0</td><td class="bad-text">+3.30</td><td>4598.3</td><td>6223.0</td><td><span class="score bad">质量差</span></td></tr>
+            </tbody>
+          </table>
+        </div>
+      </section>
+
+      <section id="env" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Environment Gap</p>
+            <h2>跨环境差异：编辑器仍会掩盖百元机红线</h2>
+          </div>
+          <p class="muted">固定同模式 / 同场景 / XZ2D / 同数量，只改变运行环境。</p>
+        </div>
+        <div class="wide-table compare-card">
+          <div class="rule"><strong>固定：</strong>RVO2 或 A Pro / XZ2D / 同场景 / 同数量。<strong>唯一变量：</strong>Editor vs Android。</div>
+          <table>
+            <thead><tr><th>配置</th><th>Editor FPS</th><th>Android FPS</th><th>Editor Sim</th><th>Android Sim</th><th>空间重叠 Editor / Android</th><th>读法</th></tr></thead>
+            <tbody>
+              <tr><td>RVO2 / FreeSwim / 300</td><td>99.3</td><td>22.5</td><td>2.08</td><td>21.36</td><td>6.0 / 16.1</td><td>Android 才暴露性能压力</td></tr>
+              <tr><td>RVO2 / BaitScramble / 300</td><td>98.5</td><td>26.5</td><td>2.25</td><td>33.41</td><td>150.9 / 122.2</td><td>鱼饵在 Android 上 sim 成本高</td></tr>
+              <tr><td>A Pro / FreeSwim / 300</td><td>100.9</td><td>30.0</td><td>1.86</td><td>14.28</td><td>5.8 / 17.9</td><td>帧率达标但质量变差</td></tr>
+              <tr><td>A Pro / BaitScramble / 300</td><td>105.6</td><td>30.1</td><td>2.01</td><td>15.62</td><td>166.1 / 235.9</td><td>鱼饵聚集在 Android 上更挤</td></tr>
+            </tbody>
+          </table>
+        </div>
+      </section>
+
+      <section id="method" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Method</p>
+            <h2>样本与统计口径</h2>
+          </div>
+        </div>
+        <div class="grid-3">
+          <div class="info-card">
+            <h3>字段</h3>
+            <p><span class="tag">averageFps</span><span class="tag">onePercentLowFps</span><span class="tag">simulationMs</span><span class="tag">spatialOverlapPairs</span></p>
+            <p class="muted">新日志还包含 averageSpeed / averageAcceleration，可用于观察鱼饵聚集是否靠剧烈运动换效果。</p>
+          </div>
+          <div class="info-card">
+            <h3>环境</h3>
+            <p><span class="tag green">Android</span><span class="tag cyan">Editor</span></p>
+            <p class="muted">Android 是主依据；Editor 只看趋势和粗略上限。</p>
+          </div>
+          <div class="info-card">
+            <h3>限制</h3>
+            <p><span class="tag orange">单次样本较多</span><span class="tag red">部分配置缺失</span></p>
+            <p class="muted">建议关键配置重复跑 3 次取中位数，尤其是 Android / BaitScramble / 300-800。</p>
+          </div>
+        </div>
+      </section>
+
+      <section id="actions" class="section">
+        <div class="section-head">
+          <div>
+            <p class="kicker">Next Actions</p>
+            <h2>建议动作</h2>
+          </div>
+        </div>
+        <div class="grid-3">
+          <div class="info-card">
+            <h3>不要直接替换</h3>
+            <p>A Pro 先不要全局替换 RVO2。它当前在鱼饵场景的质量代价太明显，尤其 500 鱼以上。</p>
+          </div>
+          <div class="info-card">
+            <h3>优先修鱼饵聚集</h3>
+            <p>给鱼饵周围加分布点、局部容量、停靠环、或重叠后处理。A Pro 的性能余量可以用来换更好的分散策略。</p>
+          </div>
+          <div class="info-card">
+            <h3>下一轮固定样本</h3>
+            <p>Android / XZ2D / BaitScramble 跑 300、500、800，每个 RVO2 与 A Pro 各 3 次，加入到达鱼饵时间和鱼饵半径内拥挤度。</p>
+          </div>
+        </div>
+        <p class="foot">生成时间：2026-06-15。数据源目录：C:\Users\13999\Desktop\logs。</p>
+      </section>
+    </main>
+  </div>
+</body>
+</html>

FishROV/Docs/avoidance-core-comparison.html

@@ -1,631 +0,0 @@
-<!doctype html>
-<html lang="zh-CN">
-<head>
-  <meta charset="utf-8">
-  <meta name="viewport" content="width=device-width, initial-scale=1">
-  <title>RVO2 避障压测方案收束</title>
-  <style>
-    :root {
-      --ink: #1d2528;
-      --muted: #667175;
-      --paper: #f5f1e8;
-      --line: #d8d0c1;
-      --green: #1f6f62;
-      --blue: #255f85;
-      --red: #a64138;
-      --gold: #a36d18;
-      --dark: #233033;
-      --soft: rgba(255, 255, 255, 0.64);
-      --shadow: 0 18px 60px rgba(35, 48, 51, 0.14);
-    }
-
-    * {
-      box-sizing: border-box;
-    }
-
-    body {
-      margin: 0;
-      color: var(--ink);
-      background:
-        linear-gradient(90deg, rgba(31, 111, 98, 0.08) 1px, transparent 1px),
-        linear-gradient(180deg, rgba(31, 111, 98, 0.06) 1px, transparent 1px),
-        var(--paper);
-      background-size: 34px 34px;
-      font-family: "Microsoft YaHei", "Noto Sans CJK SC", "PingFang SC", sans-serif;
-      line-height: 1.65;
-    }
-
-    header {
-      padding: 56px clamp(20px, 5vw, 72px) 30px;
-      border-bottom: 1px solid var(--line);
-      background: linear-gradient(135deg, rgba(255,255,255,0.72), rgba(221,232,221,0.34));
-    }
-
-    .eyebrow {
-      display: inline-flex;
-      align-items: center;
-      gap: 10px;
-      color: var(--green);
-      font-size: 13px;
-      font-weight: 700;
-      letter-spacing: 0.08em;
-      text-transform: uppercase;
-    }
-
-    .eyebrow::before {
-      content: "";
-      width: 32px;
-      height: 2px;
-      background: var(--green);
-    }
-
-    h1 {
-      max-width: 1060px;
-      margin: 18px 0 16px;
-      font-family: Georgia, "Times New Roman", serif;
-      font-size: clamp(34px, 5vw, 72px);
-      line-height: 1.02;
-      letter-spacing: 0;
-    }
-
-    .lead {
-      max-width: 980px;
-      margin: 0;
-      color: var(--dark);
-      font-size: clamp(17px, 2vw, 22px);
-    }
-
-    main {
-      padding: 34px clamp(18px, 4vw, 64px) 72px;
-    }
-
-    section {
-      max-width: 1280px;
-      margin: 0 auto 42px;
-    }
-
-    h2 {
-      margin: 0 0 18px;
-      font-size: clamp(24px, 2.8vw, 38px);
-      line-height: 1.16;
-      letter-spacing: 0;
-    }
-
-    h3 {
-      margin: 0 0 10px;
-      font-size: 18px;
-      letter-spacing: 0;
-    }
-
-    .thesis {
-      display: grid;
-      grid-template-columns: minmax(240px, 0.86fr) minmax(280px, 1.14fr);
-      gap: 22px;
-      align-items: stretch;
-    }
-
-    .verdict {
-      padding: 28px;
-      background: var(--dark);
-      color: #f8f4eb;
-      box-shadow: var(--shadow);
-    }
-
-    .verdict strong {
-      color: #f1c66d;
-    }
-
-    .shared {
-      padding: 26px;
-      background: rgba(255,255,255,0.72);
-      border: 1px solid var(--line);
-    }
-
-    .shared ul,
-    .axis-list {
-      margin: 0;
-      padding-left: 20px;
-    }
-
-    .shared li,
-    .axis-list li {
-      margin: 8px 0;
-    }
-
-    .matrix-wrap {
-      overflow-x: auto;
-      background: rgba(255,255,255,0.7);
-      border: 1px solid var(--line);
-      box-shadow: var(--shadow);
-    }
-
-    table {
-      width: 100%;
-      min-width: 1080px;
-      border-collapse: collapse;
-      background: rgba(255,255,255,0.8);
-    }
-
-    th,
-    td {
-      padding: 16px 18px;
-      border-bottom: 1px solid var(--line);
-      vertical-align: top;
-      text-align: left;
-      font-size: 15px;
-    }
-
-    th {
-      position: sticky;
-      top: 0;
-      z-index: 1;
-      color: #f8f4eb;
-      background: var(--dark);
-      font-size: 14px;
-      white-space: nowrap;
-    }
-
-    tr:last-child td {
-      border-bottom: 0;
-    }
-
-    .method {
-      font-weight: 800;
-      color: var(--dark);
-      white-space: nowrap;
-    }
-
-    .tag {
-      display: inline-block;
-      margin: 0 6px 6px 0;
-      padding: 4px 8px;
-      border: 1px solid currentColor;
-      border-radius: 999px;
-      font-size: 12px;
-      font-weight: 700;
-    }
-
-    .tag-green { color: var(--green); }
-    .tag-blue { color: var(--blue); }
-    .tag-red { color: var(--red); }
-    .tag-gold { color: var(--gold); }
-
-    .axis-grid {
-      display: grid;
-      grid-template-columns: repeat(3, minmax(240px, 1fr));
-      gap: 18px;
-    }
-
-    .axis {
-      padding: 22px;
-      background: rgba(255,255,255,0.72);
-      border-top: 6px solid var(--green);
-      box-shadow: 0 10px 30px rgba(35, 48, 51, 0.09);
-    }
-
-    .axis:nth-child(2) { border-top-color: var(--blue); }
-    .axis:nth-child(3) { border-top-color: var(--red); }
-    .axis:nth-child(4) { border-top-color: var(--gold); }
-    .axis:nth-child(5) { border-top-color: #6e5a93; }
-    .axis:nth-child(6) { border-top-color: #3f6b2f; }
-
-    .axis p {
-      margin: 0;
-      color: var(--muted);
-    }
-
-    .lane {
-      display: grid;
-      grid-template-columns: 180px 1fr;
-      gap: 14px;
-      align-items: start;
-      padding: 18px 0;
-      border-bottom: 1px solid var(--line);
-    }
-
-    .lane:last-child {
-      border-bottom: 0;
-    }
-
-    .lane-title {
-      font-weight: 800;
-      color: var(--dark);
-    }
-
-    .bars {
-      display: grid;
-      gap: 10px;
-    }
-
-    .bar {
-      display: grid;
-      grid-template-columns: 150px 1fr;
-      gap: 12px;
-      align-items: center;
-      font-size: 14px;
-    }
-
-    .bar-track {
-      height: 12px;
-      background: rgba(35, 48, 51, 0.12);
-      border-radius: 999px;
-      overflow: hidden;
-    }
-
-    .bar-fill {
-      height: 100%;
-      background: linear-gradient(90deg, var(--green), var(--gold));
-      border-radius: inherit;
-    }
-
-    .recommend {
-      display: grid;
-      grid-template-columns: repeat(3, minmax(260px, 1fr));
-      gap: 18px;
-    }
-
-    .scenario {
-      display: grid;
-      grid-template-columns: 1.02fr 0.98fr;
-      gap: 20px;
-    }
-
-    .scenario-card {
-      padding: 24px;
-      background: rgba(255,255,255,0.76);
-      border: 1px solid var(--line);
-      box-shadow: 0 10px 32px rgba(35, 48, 51, 0.08);
-    }
-
-    .flow {
-      display: grid;
-      gap: 12px;
-    }
-
-    .flow-step {
-      display: grid;
-      grid-template-columns: 34px 1fr;
-      gap: 12px;
-      align-items: start;
-    }
-
-    .num {
-      display: grid;
-      place-items: center;
-      width: 34px;
-      height: 34px;
-      color: #f8f4eb;
-      background: var(--green);
-      border-radius: 50%;
-      font-weight: 800;
-      line-height: 1;
-    }
-
-    .fit-grid {
-      display: grid;
-      grid-template-columns: repeat(4, minmax(190px, 1fr));
-      gap: 14px;
-    }
-
-    .benchmark-grid {
-      display: grid;
-      grid-template-columns: 0.9fr 1.1fr;
-      gap: 20px;
-    }
-
-    .benchmark-card {
-      padding: 24px;
-      color: #f8f4eb;
-      background: var(--dark);
-      box-shadow: var(--shadow);
-    }
-
-    .benchmark-card h3 {
-      color: #f1c66d;
-    }
-
-    .benchmark-card p {
-      margin: 0;
-    }
-
-    .test-list {
-      padding: 24px;
-      background: rgba(255,255,255,0.76);
-      border: 1px solid var(--line);
-    }
-
-    .test-list ol {
-      margin: 0;
-      padding-left: 22px;
-    }
-
-    .test-list li {
-      margin: 8px 0;
-    }
-
-    .fit {
-      padding: 18px;
-      min-height: 156px;
-      background: rgba(255,255,255,0.72);
-      border-left: 5px solid var(--blue);
-    }
-
-    .fit strong {
-      display: block;
-      margin-bottom: 6px;
-      color: var(--dark);
-    }
-
-    .fit small {
-      color: var(--muted);
-      font-size: 13px;
-    }
-
-    .choice {
-      padding: 22px;
-      background: var(--soft);
-      border: 1px solid var(--line);
-    }
-
-    .choice b {
-      color: var(--red);
-    }
-
-    .footnote {
-      color: var(--muted);
-      font-size: 13px;
-    }
-
-    @media (max-width: 900px) {
-      .thesis,
-      .scenario,
-      .benchmark-grid,
-      .axis-grid,
-      .recommend {
-        grid-template-columns: 1fr;
-      }
-
-      .fit-grid {
-        grid-template-columns: 1fr;
-      }
-
-      .lane {
-        grid-template-columns: 1fr;
-      }
-
-      .bar {
-        grid-template-columns: 120px 1fr;
-      }
-    }
-  </style>
-</head>
-<body>
-  <header>
-    <div class="eyebrow">FishROV Avoidance Benchmark</div>
-    <h1>RVO2 基准与自研版压测收束</h1>
-    <p class="lead">
-      当前阶段先忽略复杂玩法状态，只压测局部避障算法在理想和极端条件下的表现。主线收束为两项：RVO2 原版作为成熟 ORCA 基准，基于 RVO2 思想的自研版作为项目可控方案。
-    </p>
-  </header>
-
-  <main>
-    <section class="benchmark-grid" aria-labelledby="benchmark-title">
-      <div class="benchmark-card">
-        <h2 id="benchmark-title">压测结论</h2>
-        <h3>只测两条线就够</h3>
-        <p>
-          如果当前目标是忽略实际玩法状态，只验证理想与极端条件下的局部避障性能，那么建议压测对象收束为：
-          <strong>RVO2 原版</strong> 和 <strong>基于 RVO2 思想的自研版</strong>。RVO2 是稳定基准，自研版负责验证项目可控性、可调参空间和性能上限。
-        </p>
-      </div>
-
-      <div class="test-list">
-        <h2>推荐压测口径</h2>
-        <ol>
-          <li><strong>理想情况：</strong>空旷水域、均匀分布、目标方向一致或轻微交叉，验证基础吞吐和稳定收敛。</li>
-          <li><strong>中等冲突：</strong>双向穿流、圆阵向心、随机目标切换，验证重叠率、抖动和换边稳定性。</li>
-          <li><strong>极端情况：</strong>高密度聚集、初始重叠、狭窄通道、所有 agent 争同一目标点，验证退化行为。</li>
-          <li><strong>规模阶梯：</strong>100、500、1000、3000、5000 agent，记录每帧耗时、最大耗时、重叠对、平均速度损失。</li>
-          <li><strong>移动端口径：</strong>固定竖屏 1080x1920，编辑器和百元机使用相同分辨率、数量档、场景和采样窗口。</li>
-          <li><strong>公平条件：</strong>相同半径、最大速度、timeStep、目标生成、积分方式、邻居查询范围和日志采样策略。</li>
-        </ol>
-      </div>
-    </section>
-
-    <section aria-labelledby="log-title">
-      <h2 id="log-title">日志与真实性要求</h2>
-      <div class="matrix-wrap">
-        <table>
-          <thead>
-            <tr>
-              <th>环境</th>
-              <th>必须输出</th>
-              <th>不要做</th>
-              <th>推荐实现</th>
-            </tr>
-          </thead>
-          <tbody>
-            <tr>
-              <td class="method">Unity 编辑器</td>
-              <td>算法、场景、数量、分辨率、平均 FPS、1% Low、模拟耗时、最大耗时、GC、XZ/3D 重叠对。</td>
-              <td>不要每帧 `Debug.Log`，不要在热路径拼接字符串，不要每帧写文件。</td>
-              <td>内存环形缓冲 + 固定间隔采样；单轮结束后一次性导出 CSV/JSON，并在 Console 只打印摘要。</td>
-            </tr>
-            <tr>
-              <td class="method">百元机真机</td>
-              <td>同一套指标，加设备型号、Unity 图形 API、目标帧率、分辨率 1080x1920、温度/降频观察备注。</td>
-              <td>不要开详细 HUD、不要高频日志、不要连接调试器长期采样导致额外开销。</td>
-              <td>Release/Development 分开测；真机默认静默聚合，测试结束或手动按钮导出日志文件。</td>
-            </tr>
-            <tr>
-              <td class="method">性能公平性</td>
-              <td>日志开销要对两个算法一致，最好能统计采样写入耗时。</td>
-              <td>不要让某个算法额外输出调试明细，导致对比不公平。</td>
-              <td>所有算法只写结构化数值样本，统一由 benchmark runner 收集和导出。</td>
-            </tr>
-          </tbody>
-        </table>
-      </div>
-    </section>
-
-    <section class="thesis" aria-labelledby="summary-title">
-      <div class="verdict">
-        <h2 id="summary-title">一句话判断</h2>
-        <p>
-          如果当前阶段只做算法压测，<strong>RVO2 原版 + RVO2 思想自研版</strong> 就是最小有效对照组。
-          其他方案暂时不必进入压测主线，可以作为解释来源或后续扩展参考。
-        </p>
-      </div>
-
-      <div class="shared">
-        <h2>共同点</h2>
-        <ul>
-          <li>大多数方案都需要上层提供目标或期望速度，然后再做局部修正。</li>
-          <li>核心输入都离不开位置、速度、半径、邻居范围和预测时间。</li>
-          <li>都在权衡两个目标：尽量接近期望运动，同时减少未来碰撞风险。</li>
-          <li>当前压测只关心局部速度修正，不比较全局寻路、地图可达性或复杂玩法状态。</li>
-          <li>在 FishROV 的 3D/2.5D 场景里，多数方案仍需要明确如何处理 XZ 投影、高度层和真实体积冲突。</li>
-        </ul>
-      </div>
-    </section>
-
-    <section aria-labelledby="matrix-title">
-      <h2 id="matrix-title">核心矩阵</h2>
-      <div class="matrix-wrap">
-        <table>
-          <thead>
-            <tr>
-              <th>方案</th>
-              <th>核心抽象</th>
-              <th>最终如何选速度</th>
-              <th>安全性来源</th>
-              <th>最擅长</th>
-              <th>主要风险</th>
-              <th>FishROV 定位</th>
-            </tr>
-          </thead>
-          <tbody>
-            <tr>
-              <td class="method">RVO2 / ORCA</td>
-              <td><span class="tag tag-green">速度空间</span><span class="tag tag-blue">半平面约束</span><br>把每个邻居变成 ORCA 线性约束。</td>
-              <td>在所有安全半平面内，求最接近期望速度的新速度。</td>
-              <td>显式几何约束和低维优化，理论性最强。</td>
-              <td>大规模多动态 agent 的实时局部避障。</td>
-              <td>不做全局规划，2D 投影，参数不当会保守、停滞或碰撞。</td>
-              <td>最适合作为局部避障工程基准和对照组。</td>
-            </tr>
-            <tr>
-              <td class="method">HRVO</td>
-              <td><span class="tag tag-green">速度空间</span><span class="tag tag-gold">混合 VO/RVO</span><br>把 VO 与 RVO 的几何边界混合。</td>
-              <td>排除危险速度，再选择接近期望速度且更稳定的速度。</td>
-              <td>速度障碍几何，加上减少 reciprocal 振荡的边界设计。</td>
-              <td>对称对冲、互相避让时减少左右横跳。</td>
-              <td>仍是局部 2D 方法，不处理地图和动力学约束。</td>
-              <td>适合作为处理振荡问题的设计参考。</td>
-            </tr>
-            <tr>
-              <td class="method">RVO_Py_MAS</td>
-              <td><span class="tag tag-green">速度空间</span><span class="tag tag-gold">采样过滤</span><br>把动态体和静态障碍都转成速度禁区。</td>
-              <td>采样候选速度，先过滤禁区，有安全速度就选最接近期望速度的；没有安全速度则选折中。</td>
-              <td>禁区过滤加碰撞时间折中，不是严格 ORCA 求解。</td>
-              <td>学习 RVO 建模方式，快速改造原型。</td>
-              <td>没有空间索引时规模变贵，采样可能错过好速度。</td>
-              <td>适合作为可读的 RVO 原理参考，不建议直接当大规模生产方案。</td>
-            </tr>
-            <tr>
-              <td class="method">采样避障学习版</td>
-              <td><span class="tag tag-gold">候选速度</span><span class="tag tag-red">评分函数</span><br>候选速度由期望速度、偏转、减速、停止等组成。</td>
-              <td>对每个候选速度计算目标偏差、平滑、预测碰撞、当前重叠、边界等惩罚，选最低分。</td>
-              <td>启发式风险函数和权重调参，不保证无碰撞。</td>
-              <td>项目内自研、调参、验证新评分项。</td>
-              <td>权重依赖场景，高密度可能局部最优或卡住。</td>
-              <td>最适合作为 FishROV 自研避障算法的实验骨架。</td>
-            </tr>
-            <tr>
-              <td class="method">rl_rvo_nav</td>
-              <td><span class="tag tag-red">学习策略</span><span class="tag tag-green">RVO 先验</span><br>把 VO/RVO 风险编码进状态和奖励。</td>
-              <td>PPO actor 根据自身状态和 VO/RVO 序列输出速度增量。</td>
-              <td>训练经验、RVO 区域奖励、预计碰撞时间塑形。</td>
-              <td>复杂多机器人场景下学习更灵活的风险权衡。</td>
-              <td>训练成本高，泛化依赖仿真覆盖，工程可控性较弱。</td>
-              <td>建议先吸收风险表达和 reward/score 思想，不急着引入完整 RL 链路。</td>
-            </tr>
-          </tbody>
-        </table>
-      </div>
-    </section>
-
-    <section aria-labelledby="axes-title">
-      <h2 id="axes-title">真正拉开差距的六个轴</h2>
-      <div class="axis-grid">
-        <article class="axis">
-          <h3>1. 位置空间还是速度空间</h3>
-          <p>RVO2、HRVO、RVO_Py_MAS、rl_rvo_nav 都把问题转成“哪些速度会导致未来碰撞”。采样学习版虽然是评分，但最关键的预测碰撞项也已经进入速度空间。</p>
-        </article>
-        <article class="axis">
-          <h3>2. 约束、过滤还是惩罚</h3>
-          <p>ORCA 是约束求解，危险速度直接被半平面排除。RVO_Py_MAS 是采样后过滤禁区。采样学习版是给风险加惩罚，危险速度未必完全禁止。RL 是把风险变成训练信号。</p>
-        </article>
-        <article class="axis">
-          <h3>3. 谁承担避让责任</h3>
-          <p>VO 偏“我躲别人”；RVO/ORCA 假设双方共同承担；HRVO 在 reciprocal 基础上减少来回换边。采样版需要靠评分项或通行偏置来塑造行为。</p>
-        </article>
-        <article class="axis">
-          <h3>4. 是否进入主压测</h3>
-          <p>当前主压测只保留 RVO2 原版和自研版。HRVO、RVO_Py_MAS、RL-RVO 只作为思想参考，不进入编辑器选择项。</p>
-        </article>
-        <article class="axis">
-          <h3>5. 可解释性和可改性</h3>
-          <p>RVO2 理论清晰但实现复杂。自研版最容易改权重和新增评分项。RVO_Py_MAS 最适合读懂速度障碍。RL 灵活但可解释性和复现实验成本更高。</p>
-        </article>
-        <article class="axis">
-          <h3>6. 2D 与 3D 的边界</h3>
-          <p>这些方案多数原生是 2D 或平面导航。FishROV 里如果只看 XZ，会把不同高度层的对象误判为冲突，也可能漏掉真实 3D 体积风险。核心决策必须明确水平避障、高度分层和体积碰撞的关系。</p>
-        </article>
-      </div>
-    </section>
-
-    <section aria-labelledby="score-title">
-      <h2 id="score-title">工程选择倾向</h2>
-      <div class="lane">
-        <div class="lane-title">理论安全约束</div>
-        <div class="bars">
-          <div class="bar"><span>RVO2 / ORCA</span><div class="bar-track"><div class="bar-fill" style="width: 92%"></div></div></div>
-          <div class="bar"><span>HRVO</span><div class="bar-track"><div class="bar-fill" style="width: 82%"></div></div></div>
-          <div class="bar"><span>采样 / RL</span><div class="bar-track"><div class="bar-fill" style="width: 48%"></div></div></div>
-        </div>
-      </div>
-      <div class="lane">
-        <div class="lane-title">可改可调</div>
-        <div class="bars">
-          <div class="bar"><span>采样学习版</span><div class="bar-track"><div class="bar-fill" style="width: 94%"></div></div></div>
-          <div class="bar"><span>RVO_Py_MAS</span><div class="bar-track"><div class="bar-fill" style="width: 78%"></div></div></div>
-          <div class="bar"><span>RVO2</span><div class="bar-track"><div class="bar-fill" style="width: 45%"></div></div></div>
-        </div>
-      </div>
-    </section>
-
-    <section aria-labelledby="recommend-title">
-      <h2 id="recommend-title">推荐读法</h2>
-      <div class="recommend">
-        <div class="choice">
-          <h3>先压测 RVO2</h3>
-          <p>把 <b>RVO2 / ORCA</b> 当成标准基线，确认理想、拥挤、对冲、初始重叠等场景下的耗时和重叠率。</p>
-        </div>
-        <div class="choice">
-          <h3>再压测自研版</h3>
-          <p>自研版可以先吸收 RVO2 的速度空间思想，但保留可调评分、鱼群权重和性能优化入口。</p>
-        </div>
-        <div class="choice">
-          <h3>暂缓其他方案</h3>
-          <p>HRVO、RVO_Py_MAS、RL-RVO 暂时不用进编辑器主压测。先把 RVO2 和自研版两条主线测准。</p>
-        </div>
-      </div>
-    </section>
-
-    <section class="footnote">
-      <p>依据文档：rvo2-core-principles.md、unity-navmesh-core-principles.md、sampling-local-avoidance-core-principles.md、hrvo-core-principles.md、rvo-py-mas-core-principles.md、rl-rvo-nav-core-principles.md。</p>
-    </section>
-  </main>
-</body>
-</html>

FishROV/Docs/local-avoidance-benchmark-notes.md

@@ -2,19 +2,19 @@
 
 ## 目标
 
-这个 Unity 工程用于压测本地避障方案。当前阶段先收束到两条主线：
+这个 Unity 工程用于压测 FishROV 的本地避障方案。当前主线已经收束为：
 
-- `Rvo2Adapter`：成熟 RVO2 / ORCA 基准。
+- `NoAvoidanceAdapter`：无避障对照基线。
-- 项目自研 RVO2 思想版本：用于验证可控性、可调参空间和性能上限。
+- `Rvo2Adapter`：成熟 RVO2 / ORCA 本地避障基准，也是当前可用主线。
 
-当前工程入口只保留 `NoAvoidance`、`Rvo2` 和自研版。不同方案之间唯一应该变化的变量，是 `AvoidanceFrameworkKind` 选择后由 `AvoidanceAdapterRegistry` 创建的 adapter。
+此前的独立启发式避障方案已经从工程入口移除。后续复杂鱼群行为不再通过重写局部避障求解器推进，而是在 RVO2 外层增加行为参数、preferred velocity 生成、RVO2 参数映射和运动后处理。
 
 ## 共享运行入口
 
 - 场景：`Assets/Scenes/AvoidancePerfDemo.unity`
 - 入口脚本：`AvoidancePerfBootstrap`
-- Game 视口调试面板：
+- Game 视图调试面板：
-  - 避障方案选择：无避障、RVO2、自研版
+  - 避障方案选择：无避障、RVO2
   - 测试场景选择
   - 2D / 2.5D / 3D 模式
   - 观察视角切换
@@ -22,26 +22,20 @@
   - 暂停 / 继续 / 重置
   - 平均 FPS、1% Low FPS、模拟耗时、GC 分配
   - 碰撞调试：XZ 平面重叠对数、3D 空间重叠对数、最近清空间距
-- 表现对象：鱼为运行时圆锥，圆锥尖头表示朝向。
-- 共享场景：
-  - 自由巡游
-  - 对向穿流
-  - 狭窄通道
-  - 圆阵高压
 
 ## Adapter 边界
 
-框架接入应主要修改：
+框架接入主要修改：
 
 ```text
 Assets/Scripts/Benchmark/FrameworkAdapters/
 ```
 
-adapter 可以给运行时创建的对象追加框架自己的组件，但不应改变：
+adapter 可以给运行时创建的对象追加框架自己的组件，但不应该改变：
 
 - 场景生成逻辑
 - 鱼数量档
-- Game 视口调试控件
+- Game 视图调试控件
 - 指标名称
 - 相机 / 视角模式
 - spawn 参数，除非框架确实需要很小且有文档说明的兼容调整
@@ -50,7 +44,22 @@ adapter 可以给运行时创建的对象追加框架自己的组件，但不应
 
 `NoAvoidanceAdapter` 是对照基线：只按期望速度移动，不做避障。
 
-当前压测结论主线只比较 RVO2 原版和基于 RVO2 思想的自研版本。编辑器压测入口已经只保留无避障、RVO2 和自研版，避免无关选项干扰当前结论。
+`Rvo2Adapter` 是当前主线：
+
+- 直接调用 `RVO.Simulator`，避免每个 agent 每帧多次反射调用污染性能对比。
+- 每帧设置期望速度前会同步 Unity Transform 位置，避免 RVO 内部位置和 benchmark 边界夹取结果漂移。
+- RVO2-CS 是 2D 算法，2D 和 2.5D 使用 Unity `x/z` 到 RVO `x/y` 的映射；3D 模式当前会投影到 XZ 平面。
+
+## 后续扩展方向
+
+复杂鱼群场景优先做在 RVO2 外层：
+
+- 行为层：胆小、攻击性、群体倾向、追逐、逃逸、领地、避障优先级。
+- 参数映射：按鱼种或状态调整 `radius`、`neighborDist`、`maxNeighbors`、`timeHorizon`、`maxSpeed`。
+- 期望速度生成：把目标、水流、群体队形、垂直偏好合成 preferred velocity。
+- 后处理：转向速度限制、加速度限制、动画姿态和视觉平滑。
+
+只有当 RVO2 外层扩展无法覆盖真实需求时，才重新评估是否需要替换底层求解器。
 
 ## 当前压测口径
 
@@ -63,7 +72,7 @@ adapter 可以给运行时创建的对象追加框架自己的组件，但不应
 
 ## 日志策略
 
-编辑器和百元机都需要日志，但日志不能拖慢性能测试。
+编辑器和真机都需要日志，但日志不能拖慢性能测试。
 
 - 日志目录：`Application.persistentDataPath/AvoidanceBenchmarkLogs/`。
 - 编辑器下通常在用户目录的 `AppData/LocalLow/.../AvoidanceBenchmarkLogs/`。
@@ -71,13 +80,11 @@ adapter 可以给运行时创建的对象追加框架自己的组件，但不应
 - 不要每帧 `Debug.Log`。
 - 不要在热路径里拼接字符串。
 - 不要每帧写文件。
-- 不要让某个 adapter 单独输出额外明细，导致对比不公平。
 - 建议只在内存中记录结构化数值样本。
 - 当前实现固定每 `0.5s` 采样一次，先写入内存列表。
 - 单轮重置、退出或点击“导出日志”时一次性导出 CSV。
-- Console 只打印一行摘要，例如算法、场景、数量、平均 FPS、1% Low、平均模拟耗时、最大耗时、重叠对峰值。
 
-百元机真机日志建议额外记录：
+真机日志建议额外记录：
 
 - 设备型号。
 - 目标帧率。
@@ -86,20 +93,3 @@ adapter 可以给运行时创建的对象追加框架自己的组件，但不应
 - 是否 Development Build。
 - 是否连接调试器。
 - 是否出现明显发热、降频或后台干扰。
-
-## RVO2 Adapter
-
-当前已在 `Assets/ThirdParty/RVO2-CS/` 导入 `snape/RVO2-CS` 的核心 `RVOCS` 源码，并保留 Apache-2.0 许可证说明。
-
-- `Rvo2Adapter` 已从反射桥改为直接调用 `RVO.Simulator`，避免每个 agent 每帧多次反射调用污染性能对比。
-- 每帧设置期望速度前会同步 Unity Transform 位置，避免 RVO 内部位置和 benchmark 边界夹取结果漂移。
-- RVO2-CS 是 2D 算法，2D 和 2.5D 使用 Unity `x/z` 到 RVO `x/y` 的映射；3D 模式会投影到 XZ 平面。
-
-## 采样避障学习版
-
-`SamplingLocalAvoidanceAdapter` 是项目内自写的学习算法，用来和成熟框架做横向对比。
-
-- 使用空间哈希查找近邻，避免全量 O(n²) 扫描。
-- 对每个 agent 采样多组候选速度，并用目标偏差、速度平滑、预测碰撞、当前重叠、边界越界惩罚打分。
-- 它不是完整 ORCA，也不保证无碰撞；当前只作为自研候选的弱基线，表现明显差于 RVO2 是预期内结果。
-- 如果它在某些场景比 RVO2 更稳，只能说明当前启发式适合该场景，不代表算法理论上更强。

FishROV/Docs/rvo-py-mas-core-principles.md

@@ -271,7 +271,7 @@ vA_post = min(suitable_V, key=lambda v: distance(v, vA))
 要在速度空间判断：哪些速度会导致未来碰撞。
 ```
 
-它适合作为理解 RVO 思想和自研局部避障算法的入门参考；如果目标是直接上生产级大规模 agent 避障，RVO2 / ORCA 或项目内针对鱼群特征优化后的采样避障会更合适。
+它适合作为理解 RVO 思想和局部避障算法的入门参考；如果目标是直接上生产级大规模 agent 避障，当前 FishROV 主线应优先使用 RVO2 / ORCA，并把鱼群特征扩展放在 RVO2 外层行为与参数映射中。
 
 ## 参考
 

FishROV/Docs/sampling-local-avoidance-core-principles.md

@@ -1,234 +0,0 @@
-# 采样避障学习版核心原理
-
-## 一句话结论
-
-采样避障学习版的核心是 **不直接求解析最优解，而是在一批候选速度中打分选择当前最合适的速度**。它是项目内自写的学习基线，目标是可读、可改、可调参，而不是证明无碰撞。
-
-实现位置：
-
-```text
-Assets/Scripts/Benchmark/FrameworkAdapters/SamplingLocalAvoidanceAdapter.cs
-```
-
-## 为什么写这个版本
-
-RVO2/ORCA 很成熟，但内部数学和约束求解不容易快速改。
-
-NavMeshAgent 工程成熟，但它是完整导航系统，不是纯局部避障算法。
-
-学习版采样避障的目的不同：
-
-- 给你一个完全可控的 baseline。
-- 让每个评分项都能看懂、能改权重。
-- 方便测试“如果我改变风险函数，会发生什么”。
-- 为后续写更好的自研算法提供实验平台。
-
-## 基本思路
-
-每帧对每个 agent 做三步：
-
-```text
-1. 找附近邻居
-2. 生成候选速度
-3. 给候选速度打分，选择分数最低的速度
-```
-
-它不是像 RVO2 那样构造半平面约束，而是把避障变成一个启发式评分问题。
-
-## 空间哈希找邻居
-
-如果每个 agent 都和所有其他 agent 比较，复杂度是：
-
-```text
-O(n²)
-```
-
-5000 条鱼时会非常贵。
-
-所以学习版先把 agent 放进空间哈希格子：
-
-```text
-世界坐标 x/z
-  -> 根据 cellSize 计算格子坐标
-  -> 每个格子保存一批 agent
-  -> 查询时只看自己附近几个格子
-```
-
-这样大多数情况下只需要检查附近 agent，而不是全量扫描。
-
-## 候选速度
-
-候选速度主要来自：
-
-- 当前期望速度。
-- 零速度。
-- 以期望方向为基准旋转出的多个方向。
-- 不同速度倍率，例如全速和半速。
-
-可以理解为：
-
-```text
-我想往目标方向走，
-但也试试稍微偏左、偏右、减速、停下，
-看哪个风险最低。
-```
-
-当前版本每帧采样固定数量方向，优先保持实现简单和可读。
-
-## 评分函数
-
-每个候选速度都会计算一个分数：
-
-```text
-总分 =
-  目标偏差惩罚
-  + 速度平滑惩罚
-  + 预测碰撞惩罚
-  + 当前重叠惩罚
-  + 边界惩罚
-```
-
-分数越低，速度越适合当前帧。
-
-## 目标偏差惩罚
-
-agent 不能为了避障完全忘记目标。
-
-如果候选速度和期望速度差得很远，就加惩罚：
-
-```text
-goalPenalty = |candidate - preferred|²
-```
-
-这个项越大，agent 越倾向继续朝目标走。
-
-## 速度平滑惩罚
-
-如果每帧都选一个差异很大的速度，视觉上会抖。
-
-所以学习版会惩罚候选速度和上一帧速度的差异：
-
-```text
-smoothPenalty = |candidate - lastVelocity|²
-```
-
-这个项越大，运动越平滑，但反应也可能变慢。
-
-## 当前重叠惩罚
-
-如果两个 agent 当前已经重叠，必须强烈惩罚继续贴近的速度。
-
-学习版会计算：
-
-```text
-当前距离 - 两者半径和
-```
-
-如果小于 0，说明已经重叠。重叠越深，惩罚越高。
-
-这个项主要用来处理初始生成过密、或上一帧已经挤在一起的情况。
-
-## 预测碰撞惩罚
-
-只看当前距离不够。两个 agent 现在很远，但如果高速对冲，很快也会撞。
-
-学习版会估算候选速度下的最近接近时刻：
-
-```text
-relativePosition = otherPosition - selfPosition
-relativeVelocity = candidateVelocity - otherVelocity
-closestTime = clamp(-dot(relativePosition, relativeVelocity) / |relativeVelocity|², 0, timeHorizon)
-```
-
-然后计算在 `closestTime` 时两者距离是否小于安全半径。
-
-如果预测未来会撞，就加惩罚；越快撞、越深撞，惩罚越高。
-
-这一步是学习版最接近 RVO 思想的地方：它不只问“现在近不近”，而是问“这个速度会不会导致未来碰撞”。
-
-## 边界惩罚
-
-benchmark 有 arena 半径。候选速度如果把 agent 推向边界外，就加惩罚。
-
-这能减少 agent 因避障一路逃出场地边缘，然后被 `BenchmarkAgent.Integrate` 夹回边界导致的异常运动。
-
-## 每帧流程
-
-学习版每帧流程是：
-
-1. adapter 收集每个 agent 的期望速度。
-2. 构建空间哈希。
-3. 对每个 agent 生成候选速度。
-4. 对候选速度逐个打分。
-5. 选择最低分速度。
-6. 写回 `BenchmarkAgent.ApplyVelocity`。
-7. benchmark 统一积分移动。
-
-## 优点
-
-- 完全自写，许可证清晰。
-- 很容易读懂和调参。
-- 可以快速增加新的评分项。
-- 可以观察每个启发式对结果的影响。
-- 适合作为自研算法实验起点。
-
-## 局限
-
-- 不保证无碰撞。
-- 候选速度数量有限，可能错过真正好速度。
-- 权重依赖场景，换场景可能需要重新调。
-- 高密度时仍可能局部最优或互相卡住。
-- 当前实现是 CPU 主线程版本，还没有 Job/Burst 优化。
-- 它是启发式算法，不是完整 ORCA。
-
-## 和 RVO2 的区别
-
-| 维度 | RVO2 / ORCA | 采样避障学习版 |
-| --- | --- | --- |
-| 核心方法 | 速度空间半平面约束求解 | 候选速度打分 |
-| 理论性 | 更强 | 较弱 |
-| 可解释性 | 数学清晰，但实现复杂 | 每个评分项直接可读 |
-| 可调性 | 参数较少但影响大 | 权重和候选集都可改 |
-| 无碰撞保证 | 更接近可证明安全 | 不保证 |
-| 学习价值 | 学习速度障碍理论 | 学习工程化评分和调参 |
-
-## 和 NavMesh 的区别
-
-| 维度 | Unity NavMeshAgent | 采样避障学习版 |
-| --- | --- | --- |
-| 主要目标 | 地图导航和路径跟随 | 局部动态避障 |
-| 地图理解 | 有 NavMesh 可达性 | 没有地图理解 |
-| 动态群体 | 可做但不是最专门 | 专门面向局部邻居 |
-| 3D 体积 | 主要平面 | 当前也主要平面，可扩展 |
-| 可改性 | 内部黑盒较多 | 完全可改 |
-
-## 后续改进方向
-
-这个学习版后续可以逐步升级：
-
-- 增加候选速度自适应采样，而不是固定 16 方向。
-- 把邻居风险按距离、相对速度、角色类型加权。
-- 引入最大加速度和最大转向角约束。
-- 增加“右侧通行/左侧通行”偏置，减少对称死锁。
-- 增加队形或鱼群凝聚项。
-- 使用 Job System / Burst 优化空间哈希和评分。
-- 把评分函数输出到调试 HUD，观察每项惩罚占比。
-- 借鉴 RVO2，把明显危险的速度直接剔除，再对剩余速度打分。
-
-## 对自研算法的启发
-
-如果要写更好的 FishROV 避障算法，可以把学习版当成骨架：
-
-```text
-目标速度
-  -> 候选速度生成
-  -> 风险预测
-  -> 多目标评分
-  -> 速度平滑和运动约束
-```
-
-后续真正要优化的，是三个核心点：
-
-- 更准确的风险函数。
-- 更聪明的候选速度生成。
-- 更稳定的冲突解除策略。