ScaleBFM-K1: Large-Scale Humanoid Motion Tracking

Model-based motion tracking and large-scale training for Booster K1.

Adapted ScaleBFM-based motion tracking to the 22-DoF Booster K1, achieving over 95% full-clip completion on the BONES-SEED test set after 5 hours of training on a single node with 8 NVIDIA A10 GPUs.

Simulation & Hardware · K1 motion tracking in simulation and on the real robot

Implementation

  • Adapted robot dynamics, state representations, joint/body mappings, and differentiable forward kinematics for K1; verified kinematic consistency against MuJoCo.
  • Built and validated a motion-data pipeline spanning 142,220 BONES-SEED clips and 77 LAFAN1 clips, including morphology-consistent conversion, corpus packing, and train/evaluation partitioning.
  • Configured distributed training on eight GPUs and released the K1 training, evaluation, and export workflow.

Based on ScaleBFM, ScaleBFM_Plus, and SuperMimic.

Code · Back to projects