D4RL
D4RL copied to clipboard
Contact forces are zero in Ant-v2 dataset
I think this is related to this issue https://github.com/openai/gym/issues/2593
I think that the Ant-v2 datasets where collected without contact forces. For example I found no observation with non-zero component between indexes 27 to 110 in ant-medium-v2
. Is this intended or is it related to the issue linked above?
IQL agent stills learn reasonably well with this dataset (normalized return between 75 and 105) but is a bit unstable.
Is this expert score https://github.com/rail-berkeley/d4rl/blob/4aff6f8c46f62f9a57f79caa9287efefa45b6688/d4rl/gym_mujoco/init.py#L31 also derived without observing to the contact forces?
Thank you!
Config: python 3.8.12 mujoco v2.1.0 gym==0.21.0 mujoco-py==2.1.2.14 dm-control @ git+https://github.com/deepmind/dm_control@4f1a9944bf74066b1ffe982632f20e6c687d45f1 d4rl @ git+ssh://[email protected]/rail-berkeley/d4rl.git@4aff6f8c46f62f9a57f79caa9287efefa45b6688#egg=d4rl