fix: boundary termination, physics crash handler, action_scale 0.25, play script update

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

@@ -109,7 +109,7 @@ class DreamWaQCfg(Go1WalkNpEnvCfg):
         self.commands = DreamWaQCfg.Commands()
         self.control_config.stiffness = 28.0
         self.control_config.damping = 0.7
-        self.control_config.action_scale = 0.25  # 上游原值（基类默认 0.05）
+        self.control_config.action_scale = 0.25  # 上游原值
         self.noise_config.scale_joint_angle = 0.01  # 上游原值（基类默认 0.03）
         self._apply_dreamwaq()
 
@@ -530,6 +530,20 @@ class DreamWaQTask(Go1WalkTask):
 
         return obs, info
 
+    def update_terminated(self, state):
+        """基类接触终止 + hfield 边界终止（防止走出地形导致 NaN）。"""
+        state = super().update_terminated(state)
+        pose = self._body.get_pose(state.data)
+        base_xy = pose[:, :2]
+        hf = self._model.get_hfield(0)
+        b = hf.bound
+        out_of_bounds = (
+            (base_xy[:, 0] < b[0]) | (base_xy[:, 0] > b[3]) |
+            (base_xy[:, 1] < b[1]) | (base_xy[:, 1] > b[4])
+        )
+        state.terminated = state.terminated | out_of_bounds
+        return state
+
     # ── 奖励（与上游对齐，update_reward 中 × dt）──
 
     def _get_reward(self, data: mtx.SceneData, info: dict) -> dict[str, np.ndarray]:

motrix_envs/src/motrix_envs/np/env.py

@@ -183,8 +183,19 @@ class NpEnv(ABEnv):
 
     def physics_step(self):
         # motrixsim.SceneModel.step only supports single step, so we loop
-        for _ in range(self._cfg.sim_substeps):
-            self._model.step(self._state.data)
+        try:
+            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
+            n = self._state.data.shape[0]
+            self._state.terminated[:] = True
+            self._state.reward[:] = 0.0
+            if not hasattr(self, '_physics_crash_count'):
+                self._physics_crash_count = 0
+            self._physics_crash_count += 1
+            if self._physics_crash_count <= 3:
+                print(f"[WARN] physics crash #{self._physics_crash_count}: {e} — resetting {n} envs")
 
     def _prev_physics_step(self):
         state = self._state