Hi @erwincoumans
I am trying to replicate your awesome work based on the Isaacgym engine.
I have tried my best to implement everything the same as your code, including reward functions, simulation config (frictions, etc.), robot config (Kp, Kd, etc.), and training config (network structure, PPO parameters). My current difference is that the simulation runs 60hz instead of 1000hz (the decision frequency is still 30hz). However, I failed to get as good results as yours.
I posted some results I got based on A1 pace motion. And I would appreciate it if you could suggest improving the performance. I feel I missed some parts, especially some control parameters that lead to the instability of the robot.
Hi @erwincoumans I am trying to replicate your awesome work based on the Isaacgym engine. I have tried my best to implement everything the same as your code, including reward functions, simulation config (frictions, etc.), robot config (Kp, Kd, etc.), and training config (network structure, PPO parameters). My current difference is that the simulation runs 60hz instead of 1000hz (the decision frequency is still 30hz). However, I failed to get as good results as yours. I posted some results I got based on A1 pace motion. And I would appreciate it if you could suggest improving the performance. I feel I missed some parts, especially some control parameters that lead to the instability of the robot.
Thank you so much!
https://github.com/erwincoumans/motion_imitation/assets/22308667/202320b9-a26e-498b-bd6c-81761eb326b7