2025 · Robotics
GO2 Backflip Training
Reinforcement learningIsaacLabIsaacSimMujocoPPORobotics
Role
RL Engineer
Year
2025
Category
Robotics
Context
ETHRC
01.
Problem
A backflip is short and violent. Sim and hardware disagree, and a policy that only looks good in the trainer is useless.
I trained PPO in Isaac Lab with the Unitree-rl-lab stack, a curriculum, and a lot of parallel environments.
02.
Approach
03. Outcome
The policy transferred to MuJoCo and landed. I have not run it on the physical robot yet.