hopper-medium-expert-v0 about d4rl HOT 2 CLOSED

I found that hopper-medium-expert-v0 is combined with hopper-expert and hopper-mixed instead of hopper-medium. It can be confirmed with the script below:

import numpy as np
import gym, d4rl

obs_mixed  = gym.make('hopper-mixed-v0').get_dataset()['observations']
obs_expert = gym.make('hopper-expert-v0').get_dataset()['observations']
obs_medium_expert = gym.make('hopper-medium-expert-v0').get_dataset()['observations']

assert np.all(obs_medium_expert[1000000:] == obs_mixed[:]) == True
assert np.all(obs_medium_expert[:1000000] == obs_expert[:]) == True

When I checked the other two environments (halfcheetah and walker2d) is okay; medium-expert dataset is nothing but a concatenation of medium dataset and expert dataset.

It is definitely a bug and explains why BC performs so well with the environment!

from d4rl.

Hello,

Thanks for pointing this out. We have updated the dataset to include a new version as well: http://rail.eecs.berkeley.edu/datasets/offline_rl/gym_mujoco/hopper_medium_expert_v1.hdf5

We will shortly update the defaults in the code to make this change.

from d4rl.