search

roboterax / humanoid-gym

Humanoid-Gym: Reinforcement Learning for Humanoid Robot with Zero-Shot Sim2Real Transfer https://arxiv.org/abs/2404.05695
https://sites.google.com/view/humanoid-gym/
783 stars 132 forks source link

Sim 2 sim transfer #16

Open jaykorea opened 5 months ago

jaykorea commented 5 months ago

Dear author.

Recently I ‘ve found out this amazing project to enhance sim 2 real transfer. Well reviewed your code and I have some question that you guys have done for sim 2 sim framework.

First of all, I also tested sim 2 sim transfer with my own robot model(ledgged based robot), however I ve found out that It is really hard to safely control the robot on different simulation because physics engine differes I guess. Is there specific way to enhance sim 2 sim transfer(Isaac orbit to Mujoco) or any idea how to reduce the gap between different simulation.

I’m also wondering that you guys have different control term on Isaac gym environment(following the code). One is the dynamic randomization term and the other is PD control term. Which one did you use to train the policy?

Huge thanks to your open soure library again :)

wangyenjen commented 2 months ago

Hi @jaykorea,

The sim-to-real gap primarily arises from mismatches in physical parameters and dynamics between the simulation and the real environment, often due to inaccuracies in the simulation. To address this, we employ a secondary simulation for verification. Our policy is controlled within the PD control framework.