fix: NaN protection — dof nan_to_num, physics reset on crash

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

@@ -326,6 +326,13 @@ class DreamWaQTask(Go1WalkTask):
 
     def update_observation(self, state):
         data = state.data
+        # 清理物理崩溃残留的 NaN/Inf（仅在检测到时才修改，避免无谓开销）
+        dv = data.dof_vel
+        if np.any(~np.isfinite(dv)):
+            data.set_dof_vel(np.nan_to_num(np.array(dv), nan=0.0, posinf=0.0, neginf=0.0))
+        dp = data.dof_pos
+        if np.any(~np.isfinite(dp)):
+            data.set_dof_pos(np.nan_to_num(np.array(dp), nan=0.0, posinf=0.0, neginf=0.0), self._model)
         obs = self._get_obs(data, state.info)
 
         # 更新历史缓冲区

motrix_envs/src/motrix_envs/np/env.py

@@ -187,10 +187,12 @@ class NpEnv(ABEnv):
             for _ in range(self._cfg.sim_substeps):
                 self._model.step(self._state.data)
         except Exception as e:
-            # Rust panic / MotrixSim 物理崩溃 → 标记所有 env 为终止
+            # Rust panic / 物理崩溃 → 标记终止 + 清除腐蚀数据
             n = self._state.data.shape[0]
             self._state.terminated[:] = True
             self._state.reward[:] = 0.0
+            # 强制重置物理状态，防止 NaN 传播到 observations/rewards
+            self._state.data.reset(self._model)
             if not hasattr(self, '_physics_crash_count'):
                 self._physics_crash_count = 0
             self._physics_crash_count += 1