booster_mjlab: open-source sim-to-real training stack for the Booster K1 humanoid

kevin_zakka · x · 2026-09-19

A team released boostermjlab, an mjLab integration for the Booster K1 humanoid: robot model, training tasks, and an AMP motion-prior pipeline spanning simulation to the real robot.
Tasks include AMP-based velocity tracking and motion tracking (enabling acrobatics), with motion references retargeted from LAFAN1 onto the K1's parallel-ankle design; a built-in viewer lets you browse, play back, and edit clips before training.
An in-browser interactive demo runs the exported policy at 50 Hz in MuJoCo compiled to WebAssembly (14 MB), controllable via joystick, WASD, or gamepad.
On hardware, the AMP velocity policy walks after 30k training steps. The project comes from the whIRLwind team, which focuses on robot football and is recruiting students and collaborators.

More from Embodied