reinforcement learning example give errors.

I run reinforcement learning examples\rl\simple_example.ipynb and get a error at Training workflow section.
here is the error. how to resolve it?

AssertionError Traceback (most recent call last)
Cell In[7], line 24
18 vessel_kwargs = {
19 "update_kwargs": {"batch_size": 16, "repeat": 5},
20 "episode_per_iter": 50,
21 }
23 print("Training started")
---> 24 train(
25 simulator_fn=lambda position: SimpleSimulator(position, NSTEPS),
26 state_interpreter=state_interpreter,
27 action_interpreter=action_interpreter,
28 policy=policy,
29 reward=reward,
30 initial_states=cast(List[float], SimpleDataset([10.0, 50.0, 100.0])),
31 trainer_kwargs=trainer_kwargs,
32 vessel_kwargs=vessel_kwargs,
33 )
34 print("Training finished")

File e:\anaconda3\envs\qlib230510\lib\site-packages\qlib\rl\trainer\api.py:63, in train(simulator_fn, state_interpreter, action_interpreter, initial_states, policy, reward, vessel_kwargs, trainer_kwargs)
53 vessel = TrainingVessel(
54 simulator_fn=simulator_fn,
55 state_interpreter=state_interpreter,
(...)
60 **vessel_kwargs,
61 )
62 trainer = Trainer(**trainer_kwargs)
---> 63 trainer.fit(vessel)

File e:\anaconda3\envs\qlib230510\lib\site-packages\qlib\rl\trainer\trainer.py:224, in Trainer.fit(self, vessel, ckpt_path)
222 with _wrap_context(vessel.train_seed_iterator()) as iterator:
223 vector_env = self.venv_from_iterator(iterator)
--> 224 self.vessel.train(vector_env)
225 del vector_env # FIXME: Explicitly delete this object to avoid memory leak.
227 self._call_callback_hooks("on_train_end")

File e:\anaconda3\envs\qlib230510\lib\site-packages\qlib\rl\trainer\vessel.py:171, in TrainingVessel.train(self, vector_env)
168 self.policy.train()
170 with vector_env.collector_guard():
--> 171 collector = Collector(self.policy, vector_env, VectorReplayBuffer(self.buffer_size, len(vector_env)))
173 # Number of episodes collected in each training iteration can be overridden by fast dev run.
174 if self.trainer.fast_dev_run is not None:

File e:\anaconda3\envs\qlib230510\lib\site-packages\tianshou\data\collector.py:79, in Collector.init(self, policy, env, buffer, preprocess_fn, exploration_noise)
77 self._action_space = self.env.action_space
78 # avoid creating attribute outside init
---> 79 self.reset(False)

File e:\anaconda3\envs\qlib230510\lib\site-packages\tianshou\data\collector.py:130, in Collector.reset(self, reset_buffer, gym_reset_kwargs)
117 # use empty Batch for "state" so that self.data supports slicing
118 # convert empty Batch to None when passing data to policy
119 self.data = Batch(
120 obs={},
121 act={},
(...)
128 policy={}
129 )
--> 130 self.reset_env(gym_reset_kwargs)
131 if reset_buffer:
132 self.reset_buffer()

File e:\anaconda3\envs\qlib230510\lib\site-packages\tianshou\data\collector.py:146, in Collector.reset_env(self, gym_reset_kwargs)
144 """Reset all of the environments."""
145 gym_reset_kwargs = gym_reset_kwargs if gym_reset_kwargs else {}
--> 146 obs, info = self.env.reset(**gym_reset_kwargs)
147 if self.preprocess_fn:
148 processed_data = self.preprocess_fn(
149 obs=obs, info=info, env_id=np.arange(self.env_num)
150 )

File e:\anaconda3\envs\qlib230510\lib\site-packages\qlib\rl\utils\finite_env.py:233, in FiniteVectorEnv.reset(self, id)
231 id2idx = {i: k for k, i in enumerate(wrapped_id)}
232 if request_id:
--> 233 for i, o in zip(request_id, super().reset(request_id)):
234 obs[id2idx[i]] = self._postproc_env_obs(o)
236 for i, o in zip(wrapped_id, obs):

File e:\anaconda3\envs\qlib230510\lib\site-packages\tianshou\env\venvs.py:280, in BaseVectorEnv.reset(self, id, **kwargs)
277 self.workers[i].send(None, **kwargs)
278 ret_list = [self.workers[i].recv() for i in id]
--> 280 assert (
281 isinstance(ret_list[0], (tuple, list)) and len(ret_list[0]) == 2
282 and isinstance(ret_list[0][1], dict)
283 )
285 obs_list = [r[0] for r in ret_list]
287 if isinstance(obs_list[0], tuple): # type: ignore

AssertionError:

Start with reinforcement learning examples\rl\simple_example.ipynb and reproduce the failure in the Training workflow section. Trace the reset path through qlib\rl\trainer\api.py, trainer.py, vessel.py, and qlib\rl\utils\finite_env.py, then compare it with the Tianshou Collector reset assertion. Done means the example runs through training successfully or the required compatibility change is clearly documented.