小车控制
diff --git a/src/services/actService.ts b/src/services/actService.ts
index a2e1c47d..1334517e 100644
--- a/src/services/actService.ts
+++ b/src/services/actService.ts
@@ -34,24 +34,25 @@ export const actService = {
createModel() {
const imageInput = tf.input({shape: [64, 64, 3]});
- const stateInput = tf.input({shape: [14]});
+ const stateInput = tf.input({shape: [16]}); // 从 12 改为 16
const h1 = tf.layers.conv2d({filters: 16, kernelSize: 3, activation: 'relu'}).apply(imageInput) as tf.SymbolicTensor;
const h2 = tf.layers.maxPooling2d({poolSize: 2}).apply(h1) as tf.SymbolicTensor;
const h3 = tf.layers.conv2d({filters: 32, kernelSize: 3, activation: 'relu'}).apply(h2) as tf.SymbolicTensor;
const h4 = tf.layers.maxPooling2d({poolSize: 2}).apply(h3) as tf.SymbolicTensor;
const h5 = tf.layers.flatten().apply(h4) as tf.SymbolicTensor;
-
- const s1 = tf.layers.dense({units: 32, activation: 'relu'}).apply(stateInput) as tf.SymbolicTensor;
-
+
+ const s1 = tf.layers.dense({units: 64, activation: 'relu'}).apply(stateInput) as tf.SymbolicTensor;
+
const concatenated = tf.layers.concatenate().apply([h5, s1]) as tf.SymbolicTensor;
-
+
const d1 = tf.layers.dense({units: 128, activation: 'relu'}).apply(concatenated) as tf.SymbolicTensor;
- const output = tf.layers.dense({units: CHUNK_SIZE * 2, activation: 'linear'}).apply(d1) as tf.SymbolicTensor;
-
+ const d2 = tf.layers.dense({units: 64, activation: 'relu'}).apply(d1) as tf.SymbolicTensor;
+ const output = tf.layers.dense({units: CHUNK_SIZE * 2, activation: 'linear'}).apply(d2) as tf.SymbolicTensor;
+
const model = tf.model({inputs: [imageInput, stateInput], outputs: output});
model.compile({ optimizer: 'adam', loss: 'meanSquaredError' });
-
+
return model;
},
diff --git a/src/services/featureService.ts b/src/services/featureService.ts
new file mode 100644
index 00000000..9abd2db9
--- /dev/null
+++ b/src/services/featureService.ts
@@ -0,0 +1,158 @@
+import * as THREE from 'three';
+import { SimulationState } from '../types';
+
+/**
+ * 特征工程工厂 (Feature Factory)
+ * 职责：纯粹的物理和几何特征计算逻辑
+ */
+export const FeatureFactory = {
+ /**
+ * 计算相对于摄像头的目标特征
+ */
+ calculateRelativeTarget: (robotPos: {x: number, z: number, rotation: number}, targetPos: {x: number, z: number}) => {
+ // 摄像头向前偏移 0.85
+ const camX = robotPos.x + Math.sin(robotPos.rotation) * 0.85;
+ const camZ = robotPos.z + Math.cos(robotPos.rotation) * 0.85;
+
+ const dx = targetPos.x - camX;
+ const dz = targetPos.z - camZ;
+ const distance = Math.sqrt(dx * dx + dz * dz);
+
+ const angleToTarget = Math.atan2(dx, dz);
+ let angleDiff = angleToTarget - robotPos.rotation;
+
+ while (angleDiff > Math.PI) angleDiff -= Math.PI * 2;
+ while (angleDiff < -Math.PI) angleDiff += Math.PI * 2;
+
+ const FOV = Math.PI / 4; // 45度
+ const isVisible = Math.abs(angleDiff) <= FOV;
+
+ return { distance, angleDiff, isVisible, dx, dz, camX, camZ };
+ },
+
+ /**
+ * 射线投射检测遮挡
+ */
+ checkOcclusion: (camPos: {x: number, z: number}, targetPos: {x: number, z: number}, walls: THREE.Mesh[], target: THREE.Mesh) => {
+ const dx = targetPos.x - camPos.x;
+ const dz = targetPos.z - camPos.z;
+ const distance = Math.sqrt(dx * dx + dz * dz);
+
+ if (distance <= 1.0) return false;
+
+ const steps = Math.floor(distance / 0.5);
+ const stepX = dx / steps;
+ const stepZ = dz / steps;
+
+ for (let i = 1; i < steps; i++) {
+ const px = camPos.x + stepX * i;
+ const pz = camPos.z + stepZ * i;
+
+ for (const wall of walls) {
+ if (wall === target) continue;
+ const w = wall.userData.w;
+ const d = wall.userData.d;
+ const wx = wall.position.x;
+ const wz = wall.position.z;
+
+ if (px > wx - w/2 && px < wx + w/2 && pz > wz - d/2 && pz < wz + d/2) {
+ return true; // 被遮挡
+ }
+ }
+ }
+ return false;
+ }
+};
+
+// ─── 特征提取函数类型 ───
+type FeatureExtractor = (sim: any, rel: any, occ: boolean) => number;
+
+/**
+ * V1 特征映射表（14 维，兼容远程 origin/3d-front 旧版本）
+ * 语义：[x, z, angle, speed, targetDist, isColliding, 0, 0, ...]
+ */
+const V1_FEATURE_MAP: FeatureExtractor[] = [
+ /* 0 */ (sim) => sim.robotState.x,
+ /* 1 */ (sim) => sim.robotState.z,
+ /* 2 */ (sim) => sim.robotState.rotation, // 原始角度值
+ /* 3 */ (sim) => sim.robotState.velocity,
+ /* 4 */ (_sim, rel) => rel.distance, // targetDist
+ /* 5 */ (sim) => sim.isColliding ? 1.0 : 0.0,
+ // 6-13 留空（补零）
+];
+
+/**
+ * V2 特征映射表（16 维，当前版本）
+ * 语义：[x, z, sin(θ), cos(θ), speed, targetX, targetZ, ballVel, ...]
+ * 顺序极其重要：新的特征应该总是追加在末尾，以保持向后兼容。
+ */
+const V2_FEATURE_MAP: FeatureExtractor[] = [
+ /* 0 */ (sim) => sim.robotState.x,
+ /* 1 */ (sim) => sim.robotState.z,
+ /* 2 */ (sim) => Math.sin(sim.robotState.rotation),
+ /* 3 */ (sim) => Math.cos(sim.robotState.rotation),
+ /* 4 */ (sim) => sim.robotState.velocity,
+ /* 5 */ (sim) => sim.target.position.x,
+ /* 6 */ (sim) => sim.target.position.z,
+ /* 7 */ (sim) => sim.ballVelX || 0,
+ /* 8 */ (sim) => sim.ballVelZ || 0,
+ /* 9 */ (_sim, rel) => rel.distance,
+ /* 10 */ (_sim, rel) => rel.angleDiff,
+ /* 11 */ (sim) => sim.isColliding ? 1.0 : 0.0,
+ /* 12 */ (_sim, rel) => rel.isVisible ? 1.0 : 0.0,
+ /* 13 */ (_sim, _rel, occ) => occ ? 1.0 : 0.0,
+ /* 14 */ (sim) => sim.ballAccelX || 0,
+ /* 15 */ (sim) => sim.ballAccelZ || 0,
+ // --- 可以在此追加 16, 17, 18 维 ---
+];
+
+/**
+ * 版本路由表：根据维度选择对应的特征映射
+ * - 14 维 → V1（旧版兼容）
+ * - 16 维 → V2（当前版本）
+ * - 其他维度 → 默认使用 V2 并裁切/填零
+ */
+const VERSION_MAP: Record = {
+ 14: V1_FEATURE_MAP,
+ 16: V2_FEATURE_MAP,
+};
+
+/**
+ * 特征适配器 (Feature Adapter)
+ * 职责：根据模型版本/维度要求，组装特征向量
+ */
+export const FeatureAdapter = {
+ /** 已知的维度版本列表 */
+ KNOWN_DIMS: [14, 16] as readonly number[],
+
+ /**
+ * 根据指定的维度组装状态向量
+ * @param dim 目标维度 (14 → V1, 16 → V2, 其他 → 默认 V2 裁切/填零)
+ * @param simData 仿真环境实时数据
+ */
+ getFeatures: (dim: number, simData: any) => {
+ const { robotState, target, walls } = simData;
+
+ // 1. 预计算核心几何特征（供 Map 使用）
+ const rel = FeatureFactory.calculateRelativeTarget(robotState, target.position);
+ const isBlocked = FeatureFactory.checkOcclusion({ x: rel.camX, z: rel.camZ }, target.position, walls, target);
+
+ // 2. 根据维度选择对应的特征映射表
+ const featureMap = VERSION_MAP[dim] || V2_FEATURE_MAP;
+
+ if (dim > featureMap.length) {
+ console.warn(`Requested dim ${dim} exceeds feature map (${featureMap.length}). Extra dims filled with zeros.`);
+ }
+
+ // 3. 动态切片组装
+ const state = new Array(dim).fill(0);
+ for (let i = 0; i < dim; i++) {
+ if (featureMap[i]) {
+ state[i] = featureMap[i](simData, rel, isBlocked);
+ }
+ // 超出映射表范围的维度保持为 0
+ }
+
+ return state;
+ }
+};
diff --git a/src/types.ts b/src/types.ts
index d2919dcf..6cfbc446 100644
--- a/src/types.ts
+++ b/src/types.ts
@@ -13,6 +13,14 @@ export type RobotConfig = string[];
export type SceneType = 'basic' | 'living_room' | 'classroom' | 'tennis_court';
export type SceneSize = 'small' | 'medium' | 'large';
export type SceneComplexity = 'low' | 'medium' | 'high';
+export type BallMotionMode = 'fixed' | 'random' | 'mixed' | 'manual';
+
+export interface BallMotionConfig {
+ mode: BallMotionMode;
+ radius: number;
+ speed: number;
+ randomIntensity: number;
+}
export interface LogEntry {
message: string;