ProtoMotions: What NVIDIA's Humanoid Simulation Framework Contains

What the repository says it is

The README calls ProtoMotions3 a GPU-accelerated simulation and learning framework for training physically simulated digital humans and humanoid robots, and a fast prototyping platform for simulated humanoid learning tasks aimed at animation, robotics and reinforcement learning. A sibling repository, MimicKit, is named alongside it. The citation block titles the project ProtoMotions3: An Open-source Framework for Humanoid Simulation and Control, with Tessler and Jiang as joint first authors and Chechik and Fidler as joint last authors. [1]

Git tags run v2.3 on 10 October 2025, v3.0 on 12 January 2026, v3.1 on 18 February 2026 and v3.2 on 10 August 2026. The docs changelog has a release section dated 12 August 2026. The last commit on main at the time of this check was f69fdca of 16 October 2026, a documentation change. The version string in pyproject.toml reads 3.1.0 on both the v3.2 tag and main. [1]

Simulators and the limit of each

Custom simulators can be added through the base_simulator interfaces. The practical reading is that Newton and Isaac Lab are the training paths, MuJoCo is the single-environment check, and Isaac Sim is for pictures. [1] [2]

Algorithms: what the README lists and what the docs add

The README names MaskedMimic as the generative policy, ADD as a roughly fifty-line agent file, GPC for discrete latent priors with PEFT task adapters, and Kimodo, NVIDIA's text-to-motion model, for authoring data. The documentation index and experiment folders go further: AMP, ASE and PPO, a DeepMimic-style Mimic tracker, a path follower, a steering task, and a BeyondMimic L2C2 sim-to-real recipe. The experiment folders are add, amp, ase, gpc, masked_mimic, mimic, path_follower and steering. [1] [2]

MaskedMimic is the paper Unified Physics-Based Character Control Through Masked Motion Inpainting by Tessler, Guo, Nabati, Chechik and Peng, arXiv 2409.14393 of 22 September 2024, published in ACM Transactions on Graphics for SIGGRAPH Asia 2024. Its project page points to ProtoMotions as the code. [3] [4]

Robot models and datasets

The README names the SMPL body, the SOMA skeleton, the Unitree G1 and the Unitree H1_2. The robot_configs folder holds g1, h1_2, smpl, smplx, soma23 and amp. The docs index also says H1; the config folder says H1_2, which is a different robot revision. Datasets named are AMASS with over 40 hours, BONES-SEED with about 142,000 motions, PHUMA, SEED BVH and G1 CSV files, and Kimodo outputs. [1] [2]

For scale, the README gives two examples rather than a requirement: the full AMASS set in 12 hours on four A100 GPUs, and a run across 24 A100s with 13,000 motions per GPU. [1]

From simulation to a physical G1

The README states that one General Tracking Policy trained on the whole BONES-SEED dataset transfers directly to the Unitree G1 zero-shot. Deployment is a single ONNX export with the observation computation baked in, tested through the RoboJuDo framework. The repository's primary audience remains simulated characters; the G1 path is the stated route to hardware. [1]

Licences differ by artefact. Source code is Apache-2.0. NVIDIA's pretrained models are under the NVIDIA Open Model License. Third-party notices for Unitree, BeyondMimic, Isaac Lab and the SMPL body models sit in the legal folder and carry their own terms, which matter before any commercial use of a G1 policy trained here. [5]

Inconsistencies to check before relying on a version

• Version string 3.1.0 on tag v3.2: cite the commit hash in a paper or a bill of materials.
• Isaac Lab 3.0 in the README versus 12.0.0 in the docs: install from the pinned commit the README gives and record it.
• Isaac Sim 5.0 or later in the README versus 6.0 in the changelog.
• H1 in the docs versus H1_2 in the configs: the robot config decides which robot you are simulating.
• No hardware minimum stated: budget from the backend's own requirements, not from the framework.