Reinforcement Learning in Isaac Sim: Which Framework?


I am focused on using Isaac Sim for manipulator control. So far, we have successfully tried RMPFlow and are now looking into an efficient way to train reinforcement learning policies in our customized environments already designed in Isaac Sim.

I noticed there are a few options provided by Nvidia, mainly Omniverse Isaac Gym and Orbit. I also ran into a couple of Gym examples within the Isaac Sim codebase, but they cannot be extended to multiple environments.

It is, however, rather confusing to choose the best path forward. I found that Orbit has a fairly complete interface for reinforcement learning experiments, but according to their website, they are planning to continue releasing as Isaac Labs. This makes me wonder if Orbit might be deprecated in the near future. Also, a substantial number of changes in the scene seem to be necessary for using Orbit as they provide a different configuration and environment interface.

1 Like

Thank you for the question. I am wondering about this too!

Hi @a.abyaneh - As announced at GTC, Isaac lab will the unified reinforcement learning platform from Nvidia for any reinforcement learning related work.

Currently it is expected to release in the month of May, so if you can wait for few more weeks then you will have more information available.

1 Like