Open Source RoboticsRobots, software & learning

ProtoMotions

Software

Humanoid motion-learning framework with backend-specific training and pretrained tracking policies; model, dataset and robot-asset terms are separate from the source-code license. [1]

Installation & first result →Official websiteGetting started

For
Researchers working on humanoid motion imitation, retargeting and simulation transfer [1]
Hardware
Unitree G1 and H1-2 models, SMPL/SMPL-X, AMP and SOMA digital humanoids [3]
Source rights
Some licenses are unresolved. Review licenses
Unknowns
compute requirements not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
Apache-2.0, NVIDIA Open Model License (upstream model policy) · license not fully assessed for datasets · Details by part

Requirements · Research details and dates

On this page

Overview

ProtoMotions trains and evaluates humanoid motion skills across several simulation backends. A useful first step is replaying the supplied G1 deployment tracker with its matching motion file in the CPU-only MuJoCo backend. That route uses one environment and lets a reader inspect tracking before committing to GPU training infrastructure. [1] [2]

Is this for you?

Known limitations
  • Genesis is experimental/marked untested.
  • Most model cards expect the training simulator; only the documented G1 deployment tracker has the full transfer-oriented recipe.
  • IsaacLab labels conflict: badge says 3.0, detailed install says 12.0.0; both point to a pinned source revision.
  • Pretrained checkpoints, motion files and example experiments are not included in the Python package.

[2] [3] [9]

What you need

Operating system
  • Linux x86_64 explicitly required for the documented pinned IsaacLab workspace; other backends require separate checks

[2]

Languages / runtimes
  • Python

[3]

Dependencies
  • Git LFS for large assets, checkpoints and motions
  • Separate Python environment per simulator
  • Newton 1.0.0 or a selected documented backend
  • IsaacLab source revision 4ecd0b036da19ff6ad2bb4d621f886b63e9f6db8 for that route

[2]

Host computer

MuJoCo route is CPU-only and limited to one environment. Newton route specifies an NVIDIA GPU with compute capability >=5.0 and driver 545+. Training resource needs depend on backend and workload.

[2]

Installation & first result

Open the upstream getting-started guide [2] · Project documentation [1]

  • Choose a simulator and its dedicated Python environment.
  • Clone the pinned source and fetch Git LFS payloads; verify files are not pointer text.
  • For a CPU first check, use the documented MuJoCo G1 deployment tracker and matching motion file with num_envs=1.
  • Use full-eval for a finite evaluation and inspect reported tracking metrics before attempting training.

[2] [3]

Installation

Use a backend-specific environment and source/LFS checkout when checkpoints, motions or examples are needed. MuJoCo supports a CPU single-environment test.

[2]

Hardware, models & data

Supported hardware
  • Unitree G1 and H1-2 models
  • SMPL/SMPL-X, AMP and SOMA digital humanoids; physical deployment is configuration-specific

[3]

Inputs and outputs

G1 deployment tracker consumes proprioception, future reference targets and action history, producing 29 joint targets for its configured PD controller.

[5]

Evaluation

full-eval exits after motion-set evaluation; ordinary inference can run continuously. Transfer requires matching the model card and control contract.

[3] [5]

Project notes

Fetch the assets, not just the Python package

The installation guide requires Git LFS for pretrained checkpoints, motions, robot meshes and USD assets. A source archive can contain pointer text in place of those files. The Python package excludes pretrained checkpoints, motion files and example experiments, so installing the package alone is not the complete quick-start environment. Use the documented source checkout and check that the required LFS files contain actual payloads. [2]

For the MuJoCo route, use Python 3.10 or newer, the CPU PyTorch installation, the backend requirements and --num-envs 1. Pair data/pretrained_models/motion_tracker/g1-bones-deploy/last.ckpt with data/motion_for_trackers/g1_bones_seed_mini.pt; the checkpoint directory also needs its inference configuration. For a finite evaluation, add --full-eval: ordinary inference can run until interrupted. We have not downloaded the checkpoint payloads or run this evaluation. [2] [3]

Simulator choice changes the requirements

Use a separate environment for each backend. The documented Newton 1.0.0 route calls for Python 3.10 or newer, an NVIDIA GPU with compute capability at least 5.0 and driver 545 or newer; Python 3.11 avoids a documented viewer dependency build. IsaacGym uses Python 3.8. Genesis is experimental and marked untested. [2]

The IsaacLab instructions contain inconsistent version labels: the badge says 3.0, while the detailed procedure names 12.0.0 with Isaac Sim 6.0. The procedure explicitly pins IsaacLab commit 4ecd0b036da19ff6ad2bb4d621f886b63e9f6db8 and requires Python 3.12/Linux x86_64. Treat that source revision as the documented target and resolve the version-label discrepancy before substituting a packaged release. [2]

Most supplied policies are intended for their training simulator. The G1 deployment tracker has the full transfer-oriented randomization recipe, but its card still requires matching joint order, observation processing, control rate, PD gains and action scaling. It produces 29 joint targets and tracks supplied reference motion; it does not generate the reference motion itself. Real-robot use needs separate safety engineering and supervision. [5] [9]

Separate the code, model and data terms

The source code is Apache-2.0. NVIDIA-provided checkpoints and exported models use the NVIDIA Open Model License unless an individual card states an exception. Training datasets and robot assets retain their own terms; the project specifically excludes SMPL/SMPL-H assets from the Apache license. A model license does not grant dataset redistribution rights. [4] [7] [8]

This profile pins commit f69fdcaabff31846b4d2247207417b57dc788828; the checked GitHub release list is empty. It is a source-researched snapshot, not a reproduced training result. The local archive does not establish that the large LFS assets, backend installations or physical transfer work. Training time, complete hardware cost and support effort remain unmeasured. [6] [2]

What we checked

Source researched only; no installation, training, benchmark or physical hardware test performed. No affiliate links or reused media.

Updates

Upstream last updated
Oct 16, 2026 [6]
Latest release
Not established
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
commit snapshot (no GitHub releases found) / f69fdcaabff31846b4d2247207417b57dc788828 · software repository

Source files and licenses

Open learning framework with separately licensed pretrained models, external motion datasets and robot assets; deployment compatibility is model-specific.

LayerAvailability / license
softwareavailable · Apache-2.0
Repository materials only; external components retain their own terms. [7] [8]
model weightsavailable · NVIDIA Open Model License (upstream model policy)
Model cards and LFS references inspected; large checkpoint payloads not downloaded. Applies to NVIDIA-provided models unless a card states an exception. Dataset and robot-asset terms are separate. [4] [5]
datasetsavailable · License not fully assessed
Sample motion references are distributed through LFS; payloads not downloaded. AMASS/BONES-SEED and other training data retain separate terms; full dataset license assessment is unresolved. [2] [3] [4]
documentationavailable · Apache-2.0
Repository materials only; external components retain their own terms. [7]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history