Question about hopper-friction expert datasets (shift levels 0.1 and 0.5)

[ineogi2]

Hi, thank you for releasing the benchmark. I’m using it for my research.

I noticed a potential issue with the MuJoCo expert dataset:
for the hopper-friction task, the shift levels 0.1 and 0.5 appear to share the exact same dataset file/content.

Is this intended, or could it be a mistake?

Thank you.

[OffDynamicsRL]

Thank you for using ODRL in your research and sorry for the late reply. I have checked the datasets and it turns out that they are identical (BTW, the random datasets of hopper-friction 0.1/0.5 are also identical). I further check the reasons and find that the issue comes from the hopper task itself. For the hopper task, the foot geom friction is the most important for the underlying dynamics. For hopper-friction-0.1, we have

off-dynamics-rl/envs/mujoco/assets/hopper_friction_0.1.xml

Line 29 in b84ab5b

<geom friction="0.2" fromto="-0.13 0 0.1 0.26 0 0.1" name="foot_geom" size="0.06" type="capsule"/>

For hopper-friction-0.5, we have

off-dynamics-rl/envs/mujoco/assets/hopper_friction_0.5.xml

Line 29 in b84ab5b

<geom friction="1.0" fromto="-0.13 0 0.1 0.26 0 0.1" name="foot_geom" size="0.06" type="capsule"/>

It seems that for the specific hopper task, using any foot friction lie in (0, 1] incurs the same results. That being said, hopper-friction-0.1 is the same task as hopper-friction-0.5 in nature. That basically can explain that the two expert datasets are identical. Sorry for not catching that when constructing the benchmark. The results reported in ODRL paper do not seem to be affected as we did not conduct experiments on hopper-friction-0.1. As a quick inspection, other tasks like halfcheetah, ant, walker2d do not seem to suffer from this issue. Also, the friction modifications function well on other shift levels in hopper tasks. I have tagged this as a Bug and have updated the README accordingly.

Suggestions: do not use hopper-friction-0.1 task.
How to fix: the solution that come to my mind is to modify the foot friction in hopper-friction-0.1/0.5 (e.g., hopper-friction-0.1 has foot friction 1.0 and hopper-friction-0.5 has foot friction 1.5). The other three friction components remain unchanged. That could fix the issue that hopper-friction-0.1 is an invalid task but would need to use new environments and new corresponding datasets. Please let me know about your thoughts. If this seems feasible, I would update the XML files and upload new datasets.

Thank you again for raising this issue

[ineogi2]

Thank you for answering. I apprecitate your well-written descriptions of this issue.
I would rather use other datasets instead of hopper-friction-0.1 from expert type.

[OffDynamicsRL]

Thank you for the reply and sorry for any inconvenience. All the best for your ongoing work