[REQUEST] Benchmarking Prior Methods and Different Task Domains for Offline RL

[Shentao-YANG]

Hi,

Thanks for releasing such an great package!

I am a Ph.D. student working on offline RL and would like to use your datasets for my research project. When comparing your offline RL datasets with D4RL (https://arxiv.org/pdf/2004.07219.pdf), may I ask the following:

1. Do you have benchmarking results for prior methods like the Appendix Table 2 & 3 (page 15-16) of the D4RL paper?
2. Does the d4rl-pybullet dataset supports other task domain than the Gym-MuJoCo alike, such as the navigation tasks Maze2D, AntMaze, etc. in the D4RL paper (page 13). I don't think these tasks require MuJoCo engine, but it is impossible to import D4RL package without checking MuJoCo activation keys.

[takuseno]

@Shentao-YANG Thanks for making this issue! Assuming that you're mentioning d4rl-pybullet, for the first question, d3rlpy currently does not have benchmark results. But, I believe we should have it since we can produce it easily. For the second, technically it's possible to create navigation tasks with PyBullet. If there is an existing PyBullet env, we can make a new dataset for it.

[Shentao-YANG]

@takuseno Thanks for your timely reply!

For the navigation tasks, I think the D4RL package defines the environments themselves, where the physical engine therein comes from the MuJoCo engine. Theoretically, I think it is possible to make a PyBullet version of it by substituting the MuJoCo part within the code with the PyBullet counterparts.

For your reference, here is the source code for creating the AntMaze task domain in the D4RL package, and the locomotion environment they defined. The latter uses MuJoCo engine, which may be changed to the PyBullet engine.

I have very little successful experience in substituting the MuJoCo engine with PyBullet. Therefore, all I said above is at you full discretion.