ASAP method

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

1. Pre-train motion tracking policies in simulation

2. Deploy the policies in the real world and collect real-world data to train a delta (residual) action model that compensates for the
   Dynamics Mismatch

3. Then ASAP fine-tunes pre-trained policies with the delta action model integrated into the simulator

Method components

• PPO based Delta Action Learning with Asymmetric Actor-Critic framework
• Humanoid control is inherently a
  POMDP
• critic network has access to privileged information such as the global positions of the reference motion and the root linear velocity
• actor network relies solely on proprioceptive inputs and a time-phase variable
• This design not only enhances phase-based motion tracking during training but also enables a simple, phase driven motion goal for sim-to-real transfer

Robust Initialization

• Reference State Initialization from random initial states for robustness

• Termination Curriculum of Tracking Tolerance that allows flexible training by permitting larger error tolerance in the initial training stage

• Domain Randomization for pretraining

Fine tuning policy

ASAP

LeCAR-Lab • Updated 2025 Feb 22 13:55

evaluates three transfer scenarios

• IsaacGym to IsaacSim

• IsaacGym to Genesis

• IsaacGym to the real-world Unitree G1 humanoid

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

https://agile.human2humanoid.com/