Shape Mismatch Error in VecNormalize during PPO Training

[Cozokim]

Steps to reproduce: I ran the following command as suggested in the readme:

python train_policy.py ../config/mujuco_BlockedHalfCheetah/train_ppo_HCWithPos-v0.yaml -n 1 -s 123

Beginning training
100%| 200000/200000 [00:00<00:00, 176379478.55it/s]
Traceback (most recent call last):
File "/home/username/Code/ICRL-benchmarks-public/interface/train_policy.py", line 311, in
train(args)
File "/home/username/Code/ICRL-benchmarks-public/interface/train_policy.py", line 211, in train
policy_agent.learn(total_timesteps=forward_timesteps,
File "/home/username/Code/ICRL-benchmarks-public/stable_baselines3/ppo/ppo.py", line 255, in learn
return super(PPO, self).learn(
File "/home/username/Code/ICRL-benchmarks-public/stable_baselines3/common/on_policy_algorithm.py", line 223, in learn
total_timesteps, callback = self._setup_learn(
File "/home/username/Code/ICRL-benchmarks-public/stable_baselines3/common/base_class.py", line 347, in _setup_learn
self._last_obs = self.env.reset()
File "/home/username/Code/ICRL-benchmarks-public/stable_baselines3/common/vec_env/vec_normalize.py", line 157, in reset
return self.normalize_obs(obs)
File "/home/username/Code/ICRL-benchmarks-public/stable_baselines3/common/vec_env/vec_normalize.py", line 113, in normalize_obs
obs = np.clip((obs - self.obs_rms.mean) / np.sqrt(self.obs_rms.var + self.epsilon), -self.clip_obs, self.clip_obs)
ValueError: operands could not be broadcast together with shapes (1,18) (17,)
Description: I encountered this error while following the tutorial to train a policy using PPO with train_policy.py. It seems like there's a mismatch between vector shapes during normalization. Specifically, the code is trying to operate on arrays with incompatible shapes: (1,18) and (17,).

It looks like the observation shapes aren't handled correctly when normalizing observations with VecNormalize.

Request: Could anyone help with this issue? It seems like a problem with the observation vector dimensions during PPO training. Any advice on how to resolve this shape mismatch would be appreciated.

Thank you!

[Guiliang]

Hi, this issue is primarily due to a version mismatch with the 'gym' library. Please ensure that you are using version 0.21.0 of 'gym'.

[Cozokim]

Thanks for the quick response!
I had initially installed a different version due to issues with Gym 0.21.0, but I managed to figure out how to install this version, and it indeed resolved the problem.
For anyone having trouble installing Gym 0.21.0, follow this link: openai/gym#3202