ASAP

Uses real-robot data to train a correction model that closes the sim-to-real gap for agile humanoid moves.

ASAP is a whole-body humanoid control method released by Carnegie Mellon University and NVIDIA in February 2025, published at RSS 2025. Motions trained in simulation often look distorted on the real robot, because simulated physics doesn't match real motors and contact dynamics — the sim-to-real gap. ASAP works in two stages. First, it trains a motion-tracking policy in simulation on retargeted human motion data. Then it deploys that policy on the real robot to collect trajectories, and trains a delta (residual) action model that learns how much extra action the simulator needs to add to match real-robot behavior; this residual is plugged back into the simulator to fine-tune the original policy. On the Unitree G1, it outperforms common approaches like system identification and domain randomization, and the code is open-sourced.

ExampleWith ASAP, a Unitree G1 reproduced signature celebration moves from athletes like Cristiano Ronaldo, Kobe Bryant, and LeBron James, as well as forward and side jumps over a meter high.

Also called

ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills

Related

Sim-to-Real Transfer · Sim-to-Real Gap (Reality Gap) · Residual Policy · Motion Tracking · Unitree G1 · Domain Randomization

Sources

ASAP (arXiv 2502.01143)
ASAP 项目主页 (Chinese)

As of

2025-04