优化PPO充电与避障策略

扩展观测特征到157维，加入充电桩、NPC、电量安全余量、地图统计和本步清扫信息。增加低电量回充动作过滤、NPC危险区过滤，并调整奖励和终局日志以突出充电、避障和真实清扫得分。
2026-04-26 14:14:18 +08:00
parent eb3efa4df7
commit efbc612945
6 changed files with 441 additions and 52 deletions
--- a/agent_ppo/feature/definition.py
+++ b/agent_ppo/feature/definition.py
@@ -33,7 +33,7 @@ ActData = create_cls(
 # 训练样本数据：字段值为 int 时框架自动按维度处理
 SampleData = create_cls(
    "SampleData",
-    obs=Config.DIM_OF_OBSERVATION,  # 69D feature vector / 特征向量
+    obs=Config.DIM_OF_OBSERVATION,  # feature vector / 特征向量
    legal_action=Config.ACTION_NUM,  # 8D legal action mask / 合法动作掩码
    act=1,  # action index / 执行的动作
    reward=Config.VALUE_NUM,  # 1D reward / 奖励