fix: NaN chain break — rewards, std, obs, physics all protected

motrix_envs/src/motrix_envs/locomotion/go1/dreamwaq.py

@@ -419,7 +419,7 @@ class DreamWaQTask(Go1WalkTask):
                     cy = half_y - self._border - row * self._cell_size - self._cell_size / 2
                     all_origins[row, col] = [cx, cy]
             self._terrain_origins = all_origins
-            self._max_init_level = 5  # 上游原值，随机 0-5 起步
+            self._max_init_level = 3  # 先 0-3，适应后再提到 5
 
         if num_reset > 0 and self._init_done and state is not None and hasattr(state, 'info'):
             old_info = state.info
@@ -610,7 +610,8 @@ class DreamWaQTask(Go1WalkTask):
         }
         # state.reward 直接用带 dt 的值（不调 super，super 不带 dt）
         state = state.replace(reward=np.zeros(self._num_envs, dtype=np.float32))
-        for v in scaled_terms.values():
+        for k, v in scaled_terms.items():
+            v = np.nan_to_num(v, nan=0.0, posinf=0.0, neginf=0.0)
             state.reward += v
         if self._cfg.reward_config.only_positive_rewards:
             state.reward = np.maximum(state.reward, 0.0)
@@ -650,9 +651,11 @@ class DreamWaQTask(Go1WalkTask):
         return np.square(base_z - ground_level - target)
 
     def _reward_joint_power(self, data):
-        torque = np.clip(data.actuator_ctrls, -100, 100)
-        vel = np.clip(self.get_dof_vel(data), -100, 100)
-        return np.sum(np.abs(torque * vel), axis=1)
+        torque = np.nan_to_num(data.actuator_ctrls, nan=0.0)
+        vel = np.nan_to_num(self.get_dof_vel(data), nan=0.0)
+        torque = np.clip(torque, -100, 100)
+        vel = np.clip(vel, -100, 100)
+        return np.clip(np.sum(np.abs(torque * vel), axis=1), 0, 1e6)
 
     def _reward_smoothness(self, info):
         scale = self.cfg.control_config.action_scale
@@ -666,7 +669,9 @@ class DreamWaQTask(Go1WalkTask):
         return np.sum(diff, axis=1)
 
     def _reward_power_distribution(self, data):
-        torque = np.clip(data.actuator_ctrls, -100, 100)
-        vel = np.clip(self.get_dof_vel(data), -100, 100)
+        torque = np.nan_to_num(data.actuator_ctrls, nan=0.0)
+        vel = np.nan_to_num(self.get_dof_vel(data), nan=0.0)
+        torque = np.clip(torque, -100, 100)
+        vel = np.clip(vel, -100, 100)
         power = torque * vel
-        return np.var(np.abs(power), axis=1)
+        return np.nan_to_num(np.var(np.abs(power), axis=1), nan=0.0)