Hi, thank you for the code. I found that the MetaWorld envs are initialized with _freeze_rand_vec=False and _partially_observable=True, this will make the agent impossible to know the goal, which is randomly reset at the start of every episode, and not provided in the state input.
I checked this at the make_metaworld_env function in utils.py. I added this code to line 74 print(env._freeze_rand_vec,env._partially_observable) and it gives 'False True'.
Another question is, which commit did you use for your experiments? Meta-World has been updated frequently, and the version used may cause differences in results. Thank you.
Hi, thank you for the code. I found that the MetaWorld envs are initialized with _freeze_rand_vec=False and _partially_observable=True, this will make the agent impossible to know the goal, which is randomly reset at the start of every episode, and not provided in the state input.
I checked this at the
make_metaworld_envfunction in utils.py. I added this code to line 74print(env._freeze_rand_vec,env._partially_observable)and it gives 'False True'.Another question is, which commit did you use for your experiments? Meta-World has been updated frequently, and the version used may cause differences in results. Thank you.