HAF: Adapting Generalist VLAs to Humanoid Whole-Body Loco-manipulation via Hierarchical Action Flow and Spectral Latent RL

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Langzhe Gu, Chengkai Hou, Meng Li, Xinhua Wang, Jiaming Liu, Xinyuan Lv, Bowei Zhang, Shuanghao Bai, Guangrun Li, Jingyang He, Gaole Dai, Ziluo Ding, Zhiyuan Xu, Kuan Cheng, Jian Tang, Zhengping Che, Shanghang Zhang
• Venue: arXiv: cs.RO
• Published: 2026-08-17
• Collected: 2026-08-18
• URL: http://arxiv.org/abs/2608.16837v1
• PDF: https://arxiv.org/pdf/2608.16837v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

The paper introduces a framework called HAF (Humanoid Adaptation Framework) to adapt generalist VLA models for humanoid whole-body locomotion tasks. The framework is divided into two parts: HAF-VLA and HAF-Steer. HAF-VLA generates hierarchical action flows through a pretrained flow-matching VLA, while HAF-Steer is a latent reinforcement learning pipeline that refines policies for efficient real-world policy adaptation.

What the Document Says

• The paper presents the Humanoid Adaptation Framework (HAF) to apply generalist VLA models to humanoid whole-body locomotion tasks.
• HAF consists of two parts: HAF-VLA and HAF-Steer.
• HAF-VLA is a hierarchical action flow generator built on a pretrained flow-matching VLA, with three sequential stages and cross-stage KV caches to retain kinematic dependencies.
• On top of the frozen HAF-VLA, HAF-Steer is a latent offline-to-online RL pipeline that uses flow-matching invertibility and DCT-based dimensionality reduction for efficient policy refinement.
• HAF was evaluated on seven real-world humanoid locomotion tasks and outperformed vanilla single-stage VLA baselines.
• The project website is available at https://grange007.github.io/HAF

Key Concepts Mentioned

• Humanoid robots
• Vision-Language-Action (VLA) models
• Locomotion
• Whole-body loco-manipulation
• Offline behavior cloning
• Online reinforcement learning
• Flow-matching
• Kinematic dependencies
• Reinforcement Learning (RL) pipeline

Claims Made by the Document

• HAF can effectively adapt generalist VLA models to humanoid whole-body loco-manipulation.

Evidence Used

• The paper describes the design and evaluation of HAF on real-world tasks.

Methods / Approach

• Pretraining a flow-matching VLA, hierarchical action flow generation, cross-stage KV caches, latent offline-to-online RL pipeline, flow-matching invertibility, DCT-based dimensionality reduction

Limitations Stated by the Authors

The paper does not discuss potential safety risks during real-robot exploration when using online reinforcement learning.

Limitations Noted by the Agent

Agent interpretation — not from the source.

The evaluation was performed on seven specific tasks; it is unknown how HAF performs on other locomotion tasks or in different environments.

Open Questions

• What are the potential safety risks during real-robot exploration with online reinforcement learning?
• How does HAF perform on other locomotion tasks or in different environments?

Notes for Future Q&A

May include interpretation. Labeled where it does.

The paper focuses on improving whole-body coordination and task performance for humanoid locomotion using the proposed framework.

Raw Abstract / Excerpt

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. To address these bottlenecks, we introduce HAF (Humanoid Adaptation Framework), a two-part framework consisting of HAF-VLA and HAF-Steer that transfers off-the-shelf generalist VLA foundation models to humanoid whole-body loco-manipulation. HAF-VLA is a hierarchical action-flow generator built on a pretrained flow-matching VLA. It splits full-body action denoising into three sequential stages with stage embeddings and cross-stage KV caches that retain kinematic dependencies, avoiding incoherent whole-body actions from one-shot generation. On top of the frozen HAF-VLA, HAF-Steer is a latent offline-to-online RL pipeline that leverages flow-matching invertibility and DCT-based dimensionality reduction to restrict RL optimization to a compact noise subspace and train a regularized SAC policy. This avoids updating the large VLA backbone and enables efficient real-world policy refinement. Evaluated on seven real-world humanoid loco-manipulation tasks, HAF surpasses vanilla single-stage VLA baselines and improves whole-body coordination and task performance. Project website: https://grange007.github.io/HAF .