ProtoMotions (NVIDIA GPU-accelerated humanoid simulation & learning framework)

NVIDIA's open-source framework for training digital humans and humanoid robots to move using GPU physics simulation.

ProtoMotions is an open-source framework from NVIDIA Research (NVlabs); its GitHub repository was created in September 2024 and it is now on its third version, ProtoMotions3, released under the Apache 2.0 license, with authors including Chen Tessler and Xue Bin Peng. It modularizes the whole pipeline of “training a humanoid to move inside physics simulation”: the backend can be swapped between Isaac Gym, Isaac Lab, Newton, and MuJoCo, and it supports the SMPL human body model as well as robots like Unitree's H1_2 and G1. It comes with built-in motion retargeting (transferring human motion capture onto a robot's skeleton), motion tracking, MaskedMimic generative control, and terrain locomotion tasks. The README reports that training on the AMASS dataset (over 40 hours of human motion) takes about 12 hours on 4 A100 GPUs; a general-purpose tracking policy trained on roughly 142,000 BONES-SEED motions can be deployed zero-shot to a real G1 robot just by exporting an ONNX model. It also supports one-command simulator swapping for sim-to-sim testing.

ExampleRetarget a human running-and-jumping motion from AMASS onto a Unitree G1 in one step, train a tracking policy in Isaac Gym, validate it in MuJoCo, then export an ONNX model and deploy it to the real robot.

Also called

ProtoMotions3

Related

Motion Tracking · MaskedMimic · Perpetual Humanoid Control · AMASS (Archive of Motion Capture as Surface Shapes) · Motion Retargeting · NVIDIA Isaac Lab

Sources

NVlabs/ProtoMotions GitHub（ProtoMotions3）
ProtoMotions 文档 (Chinese)

As of

2026-09