Hi,
Thank you for open-sourcing this benchmark and this is very cool stuff. I believe datasets for off-dynamics environments can be valuable for the RL community. However, after digging into the paper, I found that the size of the provided datasets is quite limited (e.g. 5000 transitions for locomotion tasks with friction/gravity shift), which may be enough for off-dynamics RL but not for other tasks, e.g. meta RL where we need to perform policy optimization within each task settings. Would it be possible to provide datasets with a larger volume (comparable to the ones in D4RL) for tasks with friction shifts and gravity shifts? I believe this will significantly extend the applicability of ODRL to broader scenarios, like offline meta RL.
Thanks again for this amazing repo, and if there is anything I can clarify, please do not hesitate to contact me.
Hi,
Thank you for open-sourcing this benchmark and this is very cool stuff. I believe datasets for off-dynamics environments can be valuable for the RL community. However, after digging into the paper, I found that the size of the provided datasets is quite limited (e.g. 5000 transitions for locomotion tasks with friction/gravity shift), which may be enough for off-dynamics RL but not for other tasks, e.g. meta RL where we need to perform policy optimization within each task settings. Would it be possible to provide datasets with a larger volume (comparable to the ones in D4RL) for tasks with friction shifts and gravity shifts? I believe this will significantly extend the applicability of ODRL to broader scenarios, like offline meta RL.
Thanks again for this amazing repo, and if there is anything I can clarify, please do not hesitate to contact me.