What is scalebfm?

ScaleBFM is the official code release for a research paper called Scaling Behavior Foundation Model for Humanoid Robots. A behavior foundation model in this context is a machine learning model trained to produce general movement and behavior for a robot, rather than being built for one narrow task. As of this README, the repository is still in the process of being released. The author explains that packaging the full codebase took longer than expected, and that code and resources for motion retargeting, training, deployment, and related components will be added gradually over the following week, with most of it expected to be available by the end of July 2026. The author is personally reviewing and organizing the code before publishing it, and asks readers to check back for updates or open an issue if they run into problems. The one component already available is called ScaleRetarget, a motion retargeting toolkit. Its job is to take human motion capture data, in common formats used by motion capture datasets, and convert it into joint trajectories that a physical humanoid robot can follow. Right now it ships with a complete configuration for the Unitree G1 humanoid robot, and the repository includes short animated demos showing motion from datasets called AMASS, FineDance, and Mixamo retargeted onto the G1 model. Full documentation for ScaleRetarget, including supported datasets, environment setup, the retargeting process itself, dataset preparation, and visualization, lives in its own README inside the ScaleRetarget folder. Since this is a paper implementation still mid-release, readers interested in the training or deployment side of ScaleBFM itself should expect that code to arrive later, not in the current snapshot of the repository.