config.txt

 # 1. 性能奖励
        r_speed = 0 * vx
        r_accel = 1.0 * ax
        # 稍微降低努力奖权重 (4.0 -> 3.0)，减少过度激进
        r_effort = -2 * (np.mean(current_torque) / self.max_torque)
        
# 2. 稳定性惩罚
        
        # 左右一致性
        diff_front = abs(current_torque[0] - current_torque[1])
        diff_rear = abs(current_torque[2] - current_torque[3])
        r_consistency = -50.0 * ((diff_front + diff_rear) / self.max_torque)
        
        # 横摆角速度
        r_yaw = -2.0 * abs(yaw_rate)
        
# 3. [新增] 动作平滑性惩罚 (Action Smoothness)
        # 计算当前扭矩与上一时刻扭矩的差值绝对值
        # 假设最大跳变 500Nm，归一化后是 1.0
        # 如果每步跳变 50Nm (10%)，delta=0.1，惩罚 -5.0 分
        delta_torque = np.abs(current_torque - last_torque)
        mean_delta = np.mean(delta_torque) / self.max_torque
        r_smooth = -20.0 * mean_delta 

# 4. 滑移率约束
        avg_slip_err = np.mean(np.abs(slip - self.target_slip_ratio))
        r_slip = -10.0 * avg_slip_err
        
        if np.any(np.abs(slip) > 0.8):
            r_slip -= 20.0
            
        total_reward = r_speed + r_accel + r_effort + r_consistency + r_yaw + r_slip + r_smooth

        return total_reward