SYS: ONLINELOC: WORCESTER, MANODE: GR0X-PIUTC
[ACTIVE]

RL Locomotion Training

An Isaac Lab + PPO training pipeline for GR0X locomotion, validated against the Unitree G1 29-DOF — now producing working locomotion policies for the adaptive lower body.

  • Isaac Lab
  • Reinforcement Learning
  • PPO
  • Python
  • Simulation

Problem

A learned locomotion policy is the plan for making GR0X walk, and a policy is only as good as the training pipeline behind it. Before training on the robot’s own model, the pipeline itself had to be proven correct.

My role

I set up and debugged the training pipeline end to end.

Approach

  • Built on Isaac Lab and unitree_rl_lab with PPO.
  • Used the Unitree G1 (29 DOF) as a known-good reference to validate the pipeline before switching to the GR0X model.
  • Tracked down and fixed multiple pipeline issues, including configuration mismatches with rsl_rl and observation divergence between training and evaluation.

Results

  • Working RL-based locomotion policies for the GR0X adaptive lower body.
  • In progress: preparing sim-to-real transfer.