Problems when simulating Pupper V2 in NVIDIA Isaac Gym

[Francis Russell]

Hello everyone. I have tried to simulate Pupper V2 in Isaac Gym, aimed at learning reinforcement learning based controller for the robot. I built on top of IsaacGymEnvs's Anymal example, using URDF files from https://github.com/G-Levine/puppersim/ . However after the simulation started, the pupper just fell down and had shown no tendency to improve after many learning steps. Could anyone help me inspecting my code(https://github.com/fokx/IsaacGymEnvs)? Thanks a lot!

[Florian Golemo]

Hey Francis,

We did some experimenting and found that the URDF model as-is doesn't work well because of the low masses of the legs.

You can have a look at https://leggedrobotics.github.io/legged_gym/ which has a well-working implementation for the A1 which is slightly larger and heavier.

We ended up using the masses of the A1 in our experiments but if someone has a better idea or knows how to get the sim stable for the Pupper, I'd also be interested.

Best,

Flo

[Francis Russell]

Thanks. I have managed to enable Pupper V2 to walk using Anymal's PPO settings in IsaacGymEnvs. I figured out that the reason that the robot fell is that the 'effort' and 'velocity' value in URDF is set to zero and changed them to be the same as Anymal(see this commit: https://github.com/fokx/IsaacGymEnvs/commit/3e66481fa35f170884891ae6932b685a85f64e62).