ProtoMotions

NVIDIA framework for training simulated digital humans and humanoid robots on GPUs

License

Apache-2.0

Overview

ProtoMotions is a GPU-accelerated simulation and learning framework for training physically simulated digital humans and humanoid robots, developed by NVIDIA Research Labs. It provides a fast prototyping platform for simulated humanoid learning tasks across animation, robotics, and reinforcement learning.

The framework enables large-scale motion learning — training a fully physically simulated character to learn motion skills from the entire AMASS human animation dataset (40+ hours) within 12 hours on 4 A100 GPUs. It supports scalable multi-GPU training, one-command motion retargeting to any robot via PyRoki, sim-to-sim testing across physics engines (IsaacGym to Newton/MuJoCo Warp), high-fidelity rendering in IsaacSim 5.0+, motion authoring from text prompts, procedural scene generation, and generative policies like MaskedMimic.

ProtoMotions supports multiple simulator backends including NVIDIA Newton, IsaacLab, IsaacGym, and IsaacSim, and multiple robot morphologies including SMPL/SMPLX digital humans, Unitree H1, and Unitree G1.

The MaskedMimic generative policy, built on ProtoMotions, was published at SIGGRAPH Asia 2024.

ProtoMotions is a GPU-accelerated framework from NVIDIA Research for training physically simulated digital humans and humanoid robots. It serves as a unified prototyping platform that bridges the traditionally separate worlds of character animation and humanoid robotics, enabling researchers in both fields to share infrastructure, motion data, and trained policies.

The framework's standout capability is scale: it can train a fully physically simulated character on the complete AMASS motion dataset (over 40 hours of human animation data) in just 12 hours using 4 A100 GPUs. This is enabled by scalable multi-GPU training and tight integration with NVIDIA's simulation stack. ProtoMotions supports one-command motion retargeting to any robot morphology via PyRoki, making it straightforward to transfer human motion skills to robots like the Unitree H1 and G1.

ProtoMotions supports multiple simulator backends including NVIDIA Newton, IsaacLab, IsaacGym, and IsaacSim, as well as multiple body representations (SMPL, SMPLX, and custom robot URDFs). Notable features include sim-to-sim testing across physics engines, high-fidelity rendering in IsaacSim 5.0+, motion authoring from text prompts, and procedural scene generation for diverse training environments.

The framework includes MaskedMimic, a generative physics-based character controller published at SIGGRAPH Asia 2024, which enables prompt-driven control of physically simulated characters. ProtoMotions is actively maintained by NVIDIA and represents their primary open-source offering for humanoid motion research.

Why it matters

ProtoMotions bridges the gap between character animation and humanoid robotics research by providing a unified, GPU-accelerated framework from NVIDIA. Its ability to train on 40+ hours of motion data in 12 hours and retarget to any robot with one command dramatically lowers the barrier for humanoid motion learning research.

Problem it solves

Fragmented tooling for humanoid motion learning — researchers previously needed separate frameworks for animation, robotics, and RL, with manual effort to retarget motions and switch between simulators

Key takeaways

• Train on entire AMASS dataset (40+ hours) in 12 hours on 4 A100 GPUs
• One-command motion retargeting to any robot via PyRoki optimizer
• Supports IsaacGym, IsaacLab, Newton and IsaacSim simulator backends (the Genesis backend was removed in October 2026)
• Includes MaskedMimic generative policy (SIGGRAPH Asia 2024)

Who it is for

• Humanoid robotics researchers
• Character animation researchers
• Reinforcement learning practitioners
• NVIDIA GPU developers

Limitations

• Requires NVIDIA GPUs (tested on A100s)
• Large compute requirements for full-scale training (4-24 A100 GPUs)
• Newton simulator backend is experimental
• Complex dependency chain (IsaacGym, IsaacLab, AMASS, etc.)