Yifei Yao

GBC: Generalized Behavior-Cloning Framework for Whole-Body Humanoid Imitation

An open-source pathway from human motion to humanoid action: differentiable retargeting, DAgger-MMPPO, and an Isaac Lab training workflow.

GBC combines a motion-data pipeline with a differentiable inverse-kinematics network, a DAgger-MMPPO learning algorithm, and the MMTransformer architecture. The paper evaluates imitation on heterogeneous humanoids and transfer to novel motions.

The repository and documentation provide the public implementation and its release roadmap. The paper was first submitted to arXiv on August 13, 2025.

Reference

@misc{yao2025gbc,
  title = {GBC: Generalized Behavior-Cloning Framework for Whole-Body Humanoid Imitation},
  author = {Yao, Yifei and Luo, Chengyuan and Du, Jiaheng and He, Wentao and Lu, Jun-Guo},
  year = {2025},
  url = {https://arxiv.org/abs/2508.09960},
  eprint = {2508.09960},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO}
}

Metadata checked October 27, 2026. Original source ↗

← All publications