SuperMimic-Reproduced: Video Generator — What It Does and What You Need to Run It

SuperMimic-Reproduced is an open source video generator that reproduces the SuperMimic model on a stack of mjlab and MuJoCo Warp. The project is written in Python and was created as an independent implementation based on the research paper before the official code was released. The author positions it as an experiment log: it includes a journal, the full results of the BONES-SEED run, and a checkpoint.

What it does

As described by the author, the project reproduces the work of SuperMimic, a model for motion tracking of the Unitree G1 humanoid robot. The repository contains the results of the BONES-SEED experiment: the original set of 142,220 clips was filtered for suitability, after which 129,491 clips remained. They are packed into a compact fp16-format library and stored in the file bones_seed_compact.pt.

The authors note that the library takes up 6.7 GB and stays resident in VRAM. The data rate is 50 Hz; the original clips are recorded in CSV format at 120 frames per second. The project pins dependency versions: torch 2.14.1+cu130, mujoco 3.16, warp 1.18, and a specific mjlab commit 8bb2bc9.

What you need to run it

Running it requires Python 3.12 with the uv environment manager. As described by the author, the project was tested on an RTX 5090 with driver 580. The VRAM requirement is 6.7 GB; this figure is stated by the developer rather than obtained by measurement.

The code is distributed under the MIT license. It permits use, modification, and distribution of the code, including as part of other projects, provided the license text and copyright notice are retained.

The author specifically warns: if the PYTHONPATH system variable contains third-party Python packages, such as ROS, commands should be run with the prefix env -u PYTHONPATH to avoid conflicts.

Who it suits

The project will be useful to those studying the reproducibility of research results in the field of motion tracking for humanoid robots. It provides a ready-made entry point: pinned dependency versions, a prepared dataset, and a checkpoint. This saves you from having to build the environment and filter clips from scratch.

For those looking for a ready-made tool for generating video for their own tasks, the project is unlikely to be a good fit. The author explicitly calls it an experiment log rather than a product. For work with the official implementation, he refers to the HybridRobotics/supermimic repository.

SuperMimic-Reproduced is more of a lab notebook in code form than a standalone generator. It documents a specific experiment with specific library versions and a single dataset. That is enough to reproduce the result on an RTX 5090, but not for broader use.