Accuracy of policies trained in Isaac Gym Preview Release vs Isaac Sim

Dear team,

I tried to train a reinforcement learning policy using OmniIsaacGymEnvs in Isaac Sim and using IsaacGymEnvs in Isaac Gym Preview Release. I also tried this example: 3. Creating New RL Environment — Omniverse Robotics documentation. Is there a difference between the policy generated in Isaac Gym Preview Release and Isaac Sim due to simulation accuracy being different in the two tools?

Any help will be much appreciated.

Hi @bajshubh ,

Isaac Sim does use a more recent version of PhysX compared to what we’ve shipped in Isaac Gym Preview Release, but the two should be very similar. However, policies trained using the two tools could still have differences, which we have outlined in 5. Transferring Policies from Isaac Gym Preview Releases — Omniverse Robotics documentation.

Thanks,
Kelly

1 Like

Hi @kellyg, thank you very much for the response. Please correct me if I am wrong but I am assuming that Isaac Sim provides better physics accuracy. Does this effect the accuracy of the trained policies in some way?

Hi @bajshubh ,

We haven’t noticed any visible differences in terms of accuracy in the trained policies between Isaac Sim and Isaac Gym Preview. We have also done examples of transferring policies trained in Isaac Gym Preview to Isaac Sim and have observed similar behaviors with the two tools. Hope this helps!

Thanks,
Kelly

1 Like

Thank you, that answers the question.