search

awarelab / continual_world

79 stars 16 forks source link

support for meta-World v2 #2

Open optimass opened 2 years ago

optimass commented 2 years ago

Hello!

Amazing work!

I'm currently working on continual-world but using the meta-world v2 envs. I'm getting really different results. Is it on your roadmap to add support for it and rerun the experiments? I think this is important, as meta-world v1 was found to be problematic and IFRC there was some bugs in the rewards function of some tasks.

Thanks again for this work. I've been thinking for a while about doing CRL research on meta-world, your paper was a blessing.

dlpbc commented 1 year ago

I agree with the above comments. @optimass did you consider submitting a pull request with your changes to support meta-world v2 envs? I'll do that now.

Thanks.

optimass commented 1 year ago

hi @dlpbc. Actually, I've used the Sequoia library to import Continual-World w/ meta-world v2. Here's the code to my paper.

I have now a lot of experience w/ continual world in v2, which is quite a different beast than the v1's. Feel free to email me if you have questions or if you need tips. @zajaczajac @maciejwolczyk feel free to send people my way if they have inquiries about v2.

yingchengyang commented 1 year ago

Hi @optimass , I'm trying to change the code into torch and test in MetaWorld v2. But when I run sac to handle single task, the training stage is strange and the result can not converge. Are there any tips for training SAC in MetaWorld v2 ?

Thanks a lot!

optimass commented 1 year ago

v1 and v2 are too different for one to seamlessly jump from one to another. If you want to use torch and meta-world v2, again I'd suggest just to use my codebase for that. Happy to answer any question, here's my email