HAF: Adapting Generalist VLAs to Humanoid Whole-Body Loco-Manipulation

Forum topic · 小凯 · 2026-08-19

Summary

HAF (Humanoid Adaptation Framework) is a robotics paper (arXiv:2608.16837) by Langzhe Gu, Chengkai Hou, Meng Li et al. that adapts off-the-shelf generalist vision-language-action (VLA) foundation models to humanoid whole-body loco-manipulation. The framework has two parts: HAF-VLA, a hierarchical action flow generator built on a pretrained flow-matching VLA that decouples whole-body action denoising into three stages to preserve kinematic dependencies while coordinating locomotion, waist posture, and dual-arm manipulation; and HAF-Steer, a latent offline-to-online reinforcement learning pipeline that exploits flow-matching invertibility and DCT-based dimensionality reduction to restrict RL fine-tuning to a compact noise subspace, avoiding costly and risky direct tuning of the frozen VLA backbone. Evaluated on seven real-world humanoid loco-manipulation tasks, HAF outperforms naive single-stage VLA baselines.

Paper Overview

Field: Robotics Authors: Langzhe Gu, Chengkai Hou, Meng Li et al. (17 authors) Released: 2026-08-17 arXiv: 2608.16837

Background

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. Two key bottlenecks exist:

The high dimensionality and interdependence of humanoid motions make it difficult for conventional single-stage VLA architectures to coordinate locomotion, waist posture, and dual-arm manipulation.
Policies trained via offline behavior cloning can remain suboptimal at deployment, while direct online RL fine-tuning of large VLA backbones demands excessive computation and poses safety risks during real-robot exploration.

Method: HAF (Humanoid Adaptation Framework)

HAF consists of two components that transfer off-the-shelf generalist VLA foundation models to humanoid whole-body loco-manipulation:

HAF-VLA: a hierarchical action flow generator built on a pretrained flow-matching VLA. It splits whole-body action denoising into three stages, preserving kinematic dependencies and avoiding incoherent one-shot generation of whole-body actions.
HAF-Steer: a latent offline-to-online RL pipeline stacked on top of the frozen HAF-VLA. It leverages flow-matching invertibility and DCT-based dimensionality reduction to confine RL optimization to a compact noise subspace, enabling safe and efficient online refinement.

Results

Across seven real-world humanoid loco-manipulation tasks, HAF outperforms naive single-stage VLA baselines.

---

*Auto-collected on 2026-08-19.*