fix: terrain start level 0, clear ep_report after consume

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

@@ -410,7 +410,7 @@ class DreamWaQTask(Go1WalkTask):
             all_indices = np.tile(np.arange(self._num_cols), self._num_rows)
             all_origins = self._make_origins(all_levels, all_indices)
             self._terrain_origins = all_origins.reshape(self._num_rows, self._num_cols, 2)
-            self._max_init_level = 5
+            self._max_init_level = 0  # 从平地起步，靠课程升级
 
         if num_reset > 0 and self._init_done and state is not None and hasattr(state, 'info'):
             old_info = state.info

motrix_rl/src/motrix_rl/rslrl/torch/wrap_vec_env.py

@@ -127,9 +127,10 @@ class RslrlNpEnvWrap(VecEnv):
         if "time_outs" in state.info:
             extras["time_outs"] = torch.from_numpy(state.info["time_outs"]).to(self._device)
 
-        # 将 episode 各项奖励传入 TensorBoard
+        # 将 episode 各项奖励传入 TensorBoard（消费后清除，防止重复上报）
         if "ep_report" in state.info:
-            extras["episode"] = state.info["ep_report"]
+            extras["episode"] = dict(state.info["ep_report"])
+            del state.info["ep_report"]
 
         return obs, rewards, dones, extras
 

motrix_rl/src/motrix_rl/tasks/go1.py

@@ -135,7 +135,7 @@ class rslrl:
             runner.actor.class_name = (
                 "motrix_rl.rslrl.torch.models.cenet_actor:CENetActorModel")
             runner.actor.hidden_dims = [512, 256, 128]
-            runner.actor.init_noise_std = 0.5  # 降噪声先学站稳，再探索行走
+            runner.actor.init_noise_std = 1.0  # 上游原值
 
             # Critic：标准 MLPModel，输入 privileged_obs
             runner.critic.class_name = "MLPModel"