vwxyzjn / cleanrl

High-quality single file implementation of Deep Reinforcement Learning algorithms with research-friendly features (PPO, DQN, C51, DDPG, TD3, SAC, PPG)
http://docs.cleanrl.dev
Other
5.63k stars 639 forks source link

[Request] Dexterous Manipulation Benchmark #185

Closed kevinzakka closed 2 years ago

kevinzakka commented 2 years ago

Hey Costa,

Been working on some dexterous manipulation tasks and looking for feedback on how to improve the physics parameters / reward values. Was curious if you'd be down to run some experiments together :)

Here's a link to the repo: https://github.com/kevinzakka/dexterity

vwxyzjn commented 2 years ago

Hi @kevinzakka,

Thanks for reaching out! This looks really cool. Does dexterity have a gym API by any chance? I'd be happy to help.

kevinzakka commented 2 years ago

@vwxyzjn Unfortunately not, I'm trying to stay away from that API as it seems the devs like to keep making breaking changes :| I've used a dmc2gym wrapper for testing and it works fine though :)

kevinzakka commented 2 years ago

In terms of making this as accessible as possible to the community, do you think I should provide a wrapper in the package itself or is pointing to an open-source converter OK?

vwxyzjn commented 2 years ago

That's understandable. I think maybe having a converter example would be helpful (either you implement it or use dmc2gym)

I am trying some open source wrapper here but they don't seem to work so well...

https://gist.github.com/vwxyzjn/18e1685945f3724a605fcf0d5730bfa5

kevinzakka commented 2 years ago

I'll take a look this weekend, thanks @vwxyzjn !!

vwxyzjn commented 2 years ago

Hey just following up with this: anything I can help?

kevinzakka commented 2 years ago

Hey @vwxyzjn, going to close this for now -- got something up and running without a dmc2gym conversion. Slightly busy in the next couple of weeks, so I'll keep you posted if/when I try to create the benchmark policies for the tasks.

Cheers!

vwxyzjn commented 2 years ago

That sounds good! Your project looks really interesting and let me know if there is anything I can help.