ASAP

Sim-to-real framework for agile humanoid whole-body skills using delta action models

License

MIT

Overview

ASAP (Aligning Simulation and Real-World Physics) is a two-stage framework for transferring agile whole-body skills from simulation to real humanoid robots. It addresses the fundamental dynamics mismatch problem between simulated physics and real-world physics that causes trained policies to fail on real hardware.

The framework introduces a novel delta action model that explicitly captures residual physics differences between simulation and the real world. The pipeline consists of four stages: phase-based motion tracking pre-training, delta action model training, policy fine-tuning with the delta action model, and real-world deployment.

ASAP is built on HumanoidVerse, a multi-simulator framework that allows training across IsaacGym, IsaacSim, and Genesis simulators. It uses SMPL body models for motion retargeting from human motion capture data to humanoid robots, enabling the Unitree G1 to perform highly dynamic skills including kicks, jumps, dance moves, and athletic motions.

Developed at Carnegie Mellon University in collaboration with NVIDIA. Accepted at RSS 2025.

ASAP (Aligning Simulation and Real-World Physics) is a sim-to-real transfer framework developed at Carnegie Mellon University in collaboration with NVIDIA, accepted at RSS 2025. It enables the Unitree G1 humanoid robot to perform highly dynamic whole-body skills — including kicks, jumps, dance moves, and athletic recoveries — that transfer directly from simulation to real hardware with remarkable fidelity.

The core innovation is the delta action model, a learned residual that explicitly captures the physics gap between simulation and the real world. Rather than trying to make the simulator perfectly match reality, ASAP learns what's different and compensates. The training pipeline progresses through four stages: phase-based motion tracking pre-training, delta action model training, policy fine-tuning with the delta model, and real-world deployment via ONNX export.

ASAP is built on HumanoidVerse, a multi-simulator training framework supporting IsaacGym, IsaacSim, and Genesis. It uses SMPL body model retargeting to transfer human motion capture data from the AMASS dataset to the robot's morphology, enabling a wide repertoire of agile motions. The entire training pipeline runs on NVIDIA GPUs with Weights & Biases integration for experiment tracking.

The project represents one of the most impressive demonstrations of agile humanoid locomotion transferred from simulation, with the G1 robot performing movements that were previously considered too dynamic for reliable sim-to-real transfer.

Why it matters

ASAP represents a major advance in sim-to-real transfer for humanoid robots. By learning an explicit delta action model that captures the physics gap between simulation and reality, it enables humanoid robots to perform highly dynamic and agile motions that were previously impossible to transfer reliably from simulation.

Problem it solves

Bridging the dynamics gap between physics simulation and real-world robot hardware, which is the primary bottleneck preventing agile learned policies from working on real humanoid robots

Key takeaways

• Delta action model explicitly captures sim-to-real physics differences
• Enables highly agile humanoid motions including kicks, jumps, and dance moves on real hardware
• Built on HumanoidVerse supporting IsaacGym, IsaacSim, and Genesis simulators
• Uses SMPL-based motion retargeting from human motion capture to any humanoid robot

Who it is for

• Humanoid robotics researchers
• Sim-to-real transfer researchers
• Reinforcement learning practitioners
• Motion control engineers

Limitations

• Primarily demonstrated on Unitree G1
• Requires NVIDIA GPU with IsaacGym support
• Depends on AMASS motion capture dataset for training
• Real-world deployment requires ROS 2 and Unitree SDK integration