调整PPO奖励突出有效充电

agent_ppo/feature/preprocessor.py

@@ -527,8 +527,8 @@ class Preprocessor:
         prev_low_risk = max(0.0, self.recharge_low_battery_ratio - prev_battery_ratio)
         prev_low_risk /= max(self.recharge_low_battery_ratio, 1e-6)
         risk = max(self._recharge_risk_score(), prev_low_risk)
-        mode_bonus = 0.25 if self.was_recharge_mode or self.prev_low_battery else 0.0
-        return float(np.clip(1.2 + 1.1 * risk + mode_bonus, 1.2, 2.6))
+        mode_bonus = 0.4 if self.was_recharge_mode or self.prev_low_battery else 0.0
+        return float(np.clip(2.0 + 1.8 * risk + mode_bonus, 2.0, 4.2))
 
     def battery_fail_penalty(self):
         """Adaptive terminal penalty for running out of battery before max steps."""
@@ -951,13 +951,6 @@ class Preprocessor:
         # Step penalty / 时间惩罚
         step_penalty = -0.002
 
-        # Dense guidance: prefer moving toward visible dirt.
-        # 稠密引导：鼓励向视野内污渍靠近。
-        approach_reward = 0.0
-        if not self.recharge_mode and (self.last_nearest_dirt_dist < 200.0 or self.nearest_dirt_dist < 200.0):
-            dist_delta = float(np.clip(self.last_nearest_dirt_dist - self.nearest_dirt_dist, -5.0, 5.0))
-            approach_reward = 0.01 * dist_delta if dist_delta > 0 else 0.006 * dist_delta
-
         # Recharge guidance only activates when battery safety is the bottleneck.
         # 仅在低电量/回充模式下引导靠近充电桩，避免高电量蹲充电桩。
         charge_reward = 0.0
@@ -1015,7 +1008,6 @@ class Preprocessor:
 
         return (
             cleaning_reward
-            + approach_reward
             + charge_reward
             + exploration_reward
             + stuck_penalty