There are some small issues with the current Adroit and Pointmaze datasets.
Adroit datasets
For example, door-cloned-v1:
import minari
dataset = minari.load_dataset("door-cloned-v1", download=True)
ep = next(iter(dataset))
print(f"Obs shape {ep.observations.shape}")
print({k: v.shape for k, v in ep.infos.items()})
prints
Obs shape (201, 39)
{'success': (200,)}
This is because the AdroitHandDoor-v1 environment does not have the info["success"] key on the reset step.
Pointmaze datasets
The pointmaze datasets have the same issue described here, where some episodes do not have the correct number of observations.
There are some small issues with the current Adroit and Pointmaze datasets.
Adroit datasets
For example,
door-cloned-v1
:prints
This is because the
AdroitHandDoor-v1
environment does not have theinfo["success"]
key on the reset step.Pointmaze datasets
The pointmaze datasets have the same issue described here, where some episodes do not have the correct number of observations.