Paper Overview
Field: Robotics Authors: Langzhe Gu, Chengkai Hou, Meng Li et al. (17 authors) Published: 2026-08-17 arXiv: 2608.16837
Summary
Humanoid robots hold great promise as general-purpose agents in human-centered environments, yet generalist vision-language-action (VLA) foundation models are not readily applicable to humanoid whole-body loco-manipulation. The high dimensionality and interdependence of humanoid motions make it challenging for conventional single-stage VLA architectures to coordinate locomotion, waist posture, and dual-arm manipulation effectively. Moreover, policies trained through offline behavior cloning can remain suboptimal during real-world deployment. Although online reinforcement learning can refine policies through real-world interaction, directly tuning large VLA backbones demands excessive computation and may introduce safety risks during real-robot exploration.
Method
To address these bottlenecks, the authors introduce HAF (Humanoid Adaptation Framework), consisting of two components:
- HAF-VLA: A hierarchical action flow generator based on a pretrained flow-matching VLA. It denoises whole-body actions in three stages, preserving kinematic dependencies and avoiding generating incoherent whole-body actions in one shot.
- HAF-Steer: A latent offline-to-online RL pipeline built on top of the frozen HAF-VLA. It exploits the invertibility of flow matching and DCT-based dimensionality reduction to confine RL optimization to a compact noise subspace.
Results
Evaluated on seven real-world humanoid loco-manipulation tasks, HAF outperforms naive single-stage VLA baselines.
--- *Auto-collected on 2026-08-19*