ASAP

Why we included this project

ASAP is a reference implementation worth studying if you train legged or humanoid robots and want them to move with real agility. It tackles the sim-to-real gap by pairing a phase-based motion-tracking policy with a learned 'delta action' model that corrects the difference between simulated and real-world physics, then fine-tunes the policy with that correction. The repo includes complete training pipelines for both stages and a motion-retargeting workflow that adapts SMPL/AMASS human motion to any humanoid robot model. It runs on IsaacGym, IsaacSim/IsaacLab, and Genesis, so it assumes you already have a simulation stack rather than providing one. The code is organized around reproducible experiments, and the paper, project website, and demo video are all linked for context.