search

katerakelly / oyster

Implementation of Efficient Off-policy Meta-learning via Probabilistic Context Variables (PEARL)
MIT License
472 stars 125 forks source link

Half cheetah observations with self.get_body_com("torso").flat #14

Closed IraKorshunova closed 4 years ago

IraKorshunova commented 4 years ago

Hello! I'm wondering if there are any good reasons for adding _self.get_bodycom("torso").flat to the observations of half cheetah: https://github.com/katerakelly/oyster/blob/master/rlkit/envs/half_cheetah.py#L9

It seems to me that it only complicates comparisons between methods which also use gym. Thank you!

katerakelly commented 4 years ago

Nope, there is not a good reason for this! A little git archaeology shows that this difference goes back to differences between envs in rllab and gym. I'd expect the result to be the same, but I'm verifying this just to be sure.

IraKorshunova commented 4 years ago

thank you for the rllab reference! do you know the reason why they added it? I'm pretty sure that qpos without the x-coordinate and qvel is all you need.

katerakelly commented 4 years ago

I don't know why rllab chose to include it. It shouldn't matter for these experiments. Just to be sure, I just ran the cheetah-dir and cheetah-vel experiments without it and it achieves the same test returns as reported in the paper. I'll add a note in the code about this!