FlowMimic: Heterogeneous-Command Humanoid Control via Long-Context Flow Matching

Abstract

Humanoid whole-body control increasingly relies on heterogeneous body commands from teleoperation devices, motion-capture systems, kinematic motion generators, and high-level robot policies, ranging from dense mocap trajectories to sparse end-effector constraints and generated motion snippets. We introduce FlowMimic, a deployment-time dense-reference-free humanoid control framework that represents these inputs as masked short-horizon body targets and maps them directly to joint-level actions, without constructing command-specific dense full-body references at runtime. FlowMimic instantiates the sparse-command student as a long-context action-space flow policy: an XL-style Transformer parameterizes the conditional flow field over robot history and masked commands while distilling actions from a full-information tracking teacher. The resulting mask-conditioned interface supports live mocap teleoperation, sparse VR teleoperation, end-effector-only commands, and generated-motion inputs through the same deploy-time policy. We train on 95% of a 320+ hour retargeted mocap corpus and evaluate on a 16+ hour heterogeneous-command holdout set, with ablations over temporal context and command representation. Videos on the project website demonstrate live mocap and VR teleoperation, Kimodo-generated command inputs, and Unitree G1 deployment: https://sites.google.com/view/flowmimic.

Keywords: Humanoid Whole-Body Control, Heterogeneous Commands, Flow Matching, Teleoperation