Domain Randomization For Sim2Sim Transfer

[Luckykantnayak]

Hi @xbpeng, I have been trying simple domain randomization (DR) like Kp, Kd, COM, Mass and Joint friction for DeepMimic policy training, with 5-step robot state observation history and 1-step action history. The training of the policy seems to be very slow, w/o DR it used to take 8hrs for 8000 iterations, w/ DR it takes 8hrs for 800 iterations. And the returns after 800 iterations are also very low ( < 1). I want to test the sim2sim transfer performance of some policies for Go2 skills ( tried running deepmimic trained policy from IsaacGym to Newton, completely fails to track). Do you have any suggestions ?

[xbpeng]

It's hard to guess what might be wrong with your domain randomization implementation. It seems strange that the randomization will make the simulation so much slower. Something is probably wrong, but hard to guess what without diving into the details of your code.

Policies trained in IsaacGym will generally not transfer well to Newton because they have different simulator backends. The policy needs to be trained with proper domain randomization before this sim2sim transfer will be successful.

[Luckykantnayak]

Hi @xbpeng , I just tried running policies for Go2 shared by you on IsaacGym, IsaacLab and Newton, it seems to transfer well across the simulators visually, have to check the tracking errors. Are these trained with domain randomization ? I tested some policies of my own without DR trained in Isaac gym and tested in IsaacLab & Newton, it seems to transfer well visually.

[xbpeng]

That policy wasn't trained with domain randomization. IsaacGym and IsaacLab use the same underlying simulator, so they usually transfer well. But I am surprised to hear that it transferred well to Newton. But it's neat to hear that the Go2 policies seem to transfer well between different simulators!

[Luckykantnayak]

Even the humanoid spickiick policy worked well in Newton. I can share the videos.

[Luckykantnayak]

I would like to correct myself, the trained polcies transfer well for very stable lcomotions like go2 pace and doesn't transfer welll for very unstable or underactuated motion like humanoid spinkick.

Screencast.from.03-10-2026.04.39.28.AM.webm

Screencast.from.03-10-2026.04.46.36.AM.webm

[xbpeng]

thanks for sharing those! I think this is more or less in line with our expectations. Policies trained in Isaac will generally not transfer well to Newton. We will be adding domain randomization in the future, so this should help with sim-to-sim transfer.