kyry.me/blog/rl-workshop-atr-lab

← blog

workshop

Reinforcement learning workshop for the lab

A hands-on workshop for the ATR Lab on teaching a humanoid to move with reinforcement learning, from a naive reward to motion tracking, on the Booster K1.

Eight ways to teach a humanoid to move, each with a policy running live in MuJoCo, most of them on the Booster K1:

  1. naive RL: reward speed and staying alive, and watch the robot find strange movements;
  2. reward shaping: hand-written rewards that give a normal walk;
  3. a gait clock that schedules the steps;
  4. tracking a math-based reference gait;
  5. animation-guided RL;
  6. adversarial motion priors (AMP), learning a style from motion capture;
  7. motion tracking: dances and a cartwheel from one retargeted clip;
  8. foundation trackers that follow thousands of motions.

We also tried to break the policies on purpose (wrong observations, weaker motors, lower friction) to see what sim-to-real has to survive.

Slides will be added later.