提交全局参数配置与模块说明

critical/README.md

@@ -0,0 +1,67 @@
+# 基于 CARLA 的极端驾驶仿真场景生成算法研究
+## 项目简介
+
+研究并实现极端驾驶场景自动生成算法，定义 10 种典型危险驾驶场景，支持手动构建与强化学习（DQN / PPO 系列）自动生成对抗场景，输出标准 OpenSCENARIO (.xosc) 文件。
+
+## 技术栈
+
+- 仿真平台：hutb (CARLA 2.9.16)
+- 语言：Python 3.7
+- 强化学习：DQN、Attention-DQN、PPO、Smooth-PPO
+- 场景标准：ASAM OpenSCENARIO 1.2
+
+## 项目结构
+
+```
+├── config/              全局固定参数
+├── env/                 仿真交互环境 + 通用奖励
+├── utils/               工具函数（传感器/评估/几何）
+├── scenarios/           10 种危险场景定义
+├── rl_algorithms/       强化学习算法
+├── experiments/         训练/评估/对比入口
+├── osc_exporter/        OpenSCENARIO 导出      
+├── tests/               单元测试
+└── main.py              统一入口
+```
+
+## 10 种危险场景
+
+| # | 场景 | 类别 | 核心危险 |
+|---|------|------|----------|
+| 1 | 暴雨跟车 | 极端天气 | 湿滑路面 + 低能见度 |
+| 2 | 浓雾巡航 | 极端天气 | 能见度 < 20m |
+| 3 | 夜间黑暗行驶 | 极端天气 | 光照 5% |
+| 4 | 前车急刹 | 车辆对抗 | 前车 -8m/s² 急刹 |
+| 5 | 旁车加塞 | 车辆对抗 | 强行切入自车车道 |
+| 6 | 行人横穿 | 行人危险 | 人行道突然横穿 |
+| 7 | 鬼探头 | 行人危险 | 货车盲区突然冲出 |
+| 8 | 行人闯红灯 | 行人危险 | 违规横穿 |
+| 9 | 夜间行人横穿 | 多因素耦合 | 黑暗 + 行人 |
+| 10 | 雾天鬼探头 | 多因素耦合 | 浓雾 + 盲区 |
+
+## 快速开始
+
+```bash
+# 1. 启动 CARLA 仿真器
+cd D:\hutb\hutb && CarlaUE4.exe
+
+# 2. 运行项目
+python main.py
+
+# 或命令行模式
+python main.py train --algo dqn --scenario rain_storm --episodes 500
+python main.py evaluate --algo smooth_ppo --scenario emergency_brake --model models/xxx.pth
+python main.py export --scenario ghost_peek
+python main.py test --suite all
+```
+
+## 环境要求
+
+- Windows 10 / Ubuntu 18.04+
+- CARLA 0.9.16 (hutb 定制版)
+- Python 3.7
+- PyTorch + NumPy + PyYAML + matplotlib
+
+## 百度网盘链接
+- 通过网盘分享的文件：carla
+- 链接: https://pan.baidu.com/s/1PIzndyc4LrLxvKNImyH9CA?pwd=shju 提取码: shju 

critical/__init__.py

@@ -0,0 +1,47 @@
+# config/__init__.py
+# 全局配置模块统一入口 —— 仅导出公共固定参数
+
+from .carla_config import (
+    CARLA_HOST, CARLA_PORT, CARLA_TIMEOUT,
+    SYNC_MODE, SIMULATION_FPS, FIXED_DELTA_SECONDS,
+    EGO_VEHICLE_BLUEPRINT, ADV_VEHICLE_BLUEPRINT,
+    DEFAULT_SPAWN_INDEX,
+    COLLISION_THRESHOLD, MIN_SAFE_DISTANCE, VEHICLE_MAX_SPEED,
+    RGB_CAMERA_CONFIG, LIDAR_CONFIG, SENSOR_UPDATE_FPS,
+    MAX_EPISODE_STEPS, EPISODE_TIMEOUT,
+    COLLISION_PENALTY, SUCCESS_REWARD,
+    RESULTS_DIR, MODELS_DIR, LOGS_DIR,
+    SCENARIOS_OUTPUT_DIR, EVALUATION_OUTPUT_DIR,
+    COMPARISON_OUTPUT_DIR, PLOTS_OUTPUT_DIR,
+)
+
+from .dqn_config import (
+    STATE_SIZE as DQN_STATE_SIZE,
+    ACTION_SIZE as DQN_ACTION_SIZE,
+    HIDDEN_SIZES as DQN_HIDDEN_SIZES,
+    ATTENTION_HEAD_DIM, NUM_ATTENTION_HEADS, ATTENTION_DROPOUT,
+    LEARNING_RATE as DQN_LR, GAMMA as DQN_GAMMA,
+    TAU as DQN_TAU, TARGET_UPDATE_FREQ,
+    MEMORY_SIZE, BATCH_SIZE as DQN_BATCH_SIZE, MIN_REPLAY_SIZE,
+    EPSILON_START, EPSILON_MIN, EPSILON_DECAY,
+    TRAIN_EVERY_N_STEPS as DQN_TRAIN_EVERY,
+    MAX_EPISODES as DQN_MAX_EPISODES,
+    SAVE_EVERY_N_EPISODES as DQN_SAVE_EVERY,
+    LOG_EVERY_N_EPISODES as DQN_LOG_EVERY,
+)
+
+from .ppo_config import (
+    STATE_SIZE as PPO_STATE_SIZE,
+    ACTION_SIZE as PPO_ACTION_SIZE, ACTION_STD,
+    HIDDEN_SIZES as PPO_HIDDEN_SIZES,
+    ACTOR_HIDDEN, CRITIC_HIDDEN, ACTIVATION as PPO_ACTIVATION,
+    GAMMA as PPO_GAMMA, LAMBDA as PPO_LAMBDA,
+    EPS_CLIP, UPDATE_EVERY as PPO_UPDATE_EVERY,
+    UPDATE_POLICY_TIMES, BATCH_SIZE as PPO_BATCH_SIZE,
+    LR_ACTOR, LR_CRITIC,
+    SMOOTH_ENABLED, SMOOTH_EPS_LOW, SMOOTH_EPS_HIGH, SMOOTH_BETA,
+    VALUE_LOSS_COEF, ENTROPY_COEF, MAX_GRAD_NORM,
+    MAX_EPISODES as PPO_MAX_EPISODES,
+    SAVE_EVERY_N_EPISODES as PPO_SAVE_EVERY,
+    LOG_EVERY_N_EPISODES as PPO_LOG_EVERY,
+)

critical/carla_config.py

@@ -0,0 +1,64 @@
+# config/carla_config.py
+# 全局固定参数：CARLA 连接、地图、车型、传感器通用配置、全局路径
+# 按 config/CLAUDE.md 规范 —— 不写入任何场景差异化内容
+
+# ============================================================
+# 1. CARLA 连接公共参数
+# ============================================================
+CARLA_HOST = "localhost"
+CARLA_PORT = 2000
+CARLA_TIMEOUT = 30.0
+
+# 同步模式（强化学习训练必须）
+SYNC_MODE = True
+SIMULATION_FPS = 20
+FIXED_DELTA_SECONDS = 1.0 / SIMULATION_FPS
+
+# ============================================================
+# 2. 全局车辆公共参数
+# ============================================================
+EGO_VEHICLE_BLUEPRINT = "vehicle.tesla.model3"
+ADV_VEHICLE_BLUEPRINT = "vehicle.audi.a2"
+
+# 全局默认生成点索引（场景可覆盖）
+DEFAULT_SPAWN_INDEX = 0
+
+# 碰撞检测阈值 & 控制频率（所有场景共用）
+COLLISION_THRESHOLD = 1.5            # 碰撞判定距离 (m)
+MIN_SAFE_DISTANCE = 3.0              # 最小安全距离 (m)
+VEHICLE_MAX_SPEED = 80.0             # 最高速度 (km/h)
+
+# ============================================================
+# 3. 全局传感器公共参数
+# ============================================================
+RGB_CAMERA_CONFIG = {
+    "image_size_x": "640",
+    "image_size_y": "480",
+    "fov": "90.0",
+}
+
+LIDAR_CONFIG = {
+    "points_per_second": "100000",
+    "range": "50.0",
+}
+
+SENSOR_UPDATE_FPS = 20               # 传感器采集频率（全局统一）
+
+# ============================================================
+# 4. 训练全局参数（所有算法共用）
+# ============================================================
+MAX_EPISODE_STEPS = 500
+EPISODE_TIMEOUT = 60.0               # 单集超时 (s)
+COLLISION_PENALTY = -10.0            # 碰撞惩罚
+SUCCESS_REWARD = 10.0                # 安全完成奖励
+
+# ============================================================
+# 5. 全局路径
+# ============================================================
+RESULTS_DIR = "results"
+MODELS_DIR = "models"
+LOGS_DIR = "logs"
+SCENARIOS_OUTPUT_DIR = "results/scenarios"
+EVALUATION_OUTPUT_DIR = "results/evaluation"
+COMPARISON_OUTPUT_DIR = "results/comparison"
+PLOTS_OUTPUT_DIR = "results/plots"

critical/dqn_config.py

@@ -0,0 +1,51 @@
+# config/dqn_config.py
+# DQN / Attention-DQN 超参数配置
+
+# ============================================================
+# 环境接口
+# ============================================================
+STATE_SIZE = 6        # [ego_x, ego_y, adv_x, adv_y, ego_speed, distance]
+ACTION_SIZE = 4       # 0=无操作, 1=左转, 2=右转, 3=制动
+
+# ============================================================
+# 网络结构
+# ============================================================
+HIDDEN_SIZES = [256, 256, 128]   # 三层全连接隐藏层
+ACTIVATION = "relu"
+
+# ============================================================
+# Attention-DQN 专属
+# ============================================================
+ATTENTION_HEAD_DIM = 64          # 每个注意力头的维度
+NUM_ATTENTION_HEADS = 4          # 多头注意力头数
+ATTENTION_DROPOUT = 0.1
+
+# ============================================================
+# 训练超参数
+# ============================================================
+LEARNING_RATE = 1e-3
+GAMMA = 0.95                     # 折扣因子
+TAU = 0.005                      # 目标网络软更新系数
+TARGET_UPDATE_FREQ = 100         # 每 N 步硬更新目标网络
+
+# ============================================================
+# 经验回放池
+# ============================================================
+MEMORY_SIZE = 50000              # 回放池容量
+BATCH_SIZE = 128
+MIN_REPLAY_SIZE = 2000           # 开始训练前最少经验条数
+
+# ============================================================
+# 探索策略 (ε-greedy)
+# ============================================================
+EPSILON_START = 1.0
+EPSILON_MIN = 0.01
+EPSILON_DECAY = 0.995            # 每步乘以该系数
+
+# ============================================================
+# 训练控制
+# ============================================================
+TRAIN_EVERY_N_STEPS = 4          # 每 N 步执行一次学习
+MAX_EPISODES = 500
+SAVE_EVERY_N_EPISODES = 100
+LOG_EVERY_N_EPISODES = 10

critical/ppo_config.py

@@ -0,0 +1,55 @@
+# config/ppo_config.py
+# PPO / Smooth-PPO 超参数配置
+
+# ============================================================
+# 环境接口
+# ============================================================
+STATE_SIZE = 6        # [ego_x, ego_y, adv_x, adv_y, ego_speed, distance]
+ACTION_SIZE = 4       # 0=无操作, 1=左转, 2=右转, 3=制动
+ACTION_STD = 0.5      # 连续动作空间下初始标准差（离散动作为 softmax 温度）
+
+# ============================================================
+# 网络结构 (Actor-Critic)
+# ============================================================
+HIDDEN_SIZES = [256, 256]        # Actor 与 Critic 共享特征提取层
+ACTOR_HIDDEN = [128, 64]         # Actor 头部
+CRITIC_HIDDEN = [128, 64]        # Critic 头部
+ACTIVATION = "tanh"
+
+# ============================================================
+# PPO 核心超参数
+# ============================================================
+GAMMA = 0.99                     # 折扣因子
+LAMBDA = 0.95                    # GAE 平滑系数
+EPS_CLIP = 0.2                   # PPO 标准裁剪范围
+UPDATE_EVERY = 2048              # 每收集多少步经验后执行一次策略更新
+UPDATE_POLICY_TIMES = 10         # 每次更新时对同一批数据重复训练轮数
+BATCH_SIZE = 64                  # 小批量大小
+
+# ============================================================
+# 学习率
+# ============================================================
+LR_ACTOR = 1e-4
+LR_CRITIC = 1e-3
+
+# ============================================================
+# Smooth-PPO 专属参数（创新点）
+# ============================================================
+SMOOTH_ENABLED = True
+SMOOTH_EPS_LOW = 0.1             # 平滑裁剪下界
+SMOOTH_EPS_HIGH = 0.3            # 平滑裁剪上界
+SMOOTH_BETA = 0.5                # 平滑插值系数（0=标准, 1=全平滑）
+
+# ============================================================
+# 损失权重与正则化
+# ============================================================
+VALUE_LOSS_COEF = 0.5            # 价值损失权重
+ENTROPY_COEF = 0.01              # 熵正则化系数
+MAX_GRAD_NORM = 0.5              # 梯度裁剪阈值
+
+# ============================================================
+# 训练控制
+# ============================================================
+MAX_EPISODES = 500
+SAVE_EVERY_N_EPISODES = 100
+LOG_EVERY_N_EPISODES = 10