Humanoid Adaptation Framework

HAF

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

Langzhe Gu^1,2*, Chengkai Hou^1,2*, Meng Li^2*, Xinhua Wang², Jiaming Liu¹, Xinyuan Lv^2,3, Bowei Zhang^2,3, Shuanghao Bai^2,4, Guangrun Li^1,2, Jingyang He^1,2, Gaole Dai¹, Ziluo Ding², Zhiyuan Xu², Kuan Cheng¹, Jian Tang^2✉, Zhengping Che^2✉, Shanghang Zhang^1✉

¹ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University ² Beijing Innovation Center of Humanoid Robotics ³ Nankai University ⁴ Xi'an Jiaotong University

* Equal contribution ✉ Corresponding authors

Explore the framework Read the paper

HAF-VLAHAF-Steer

Explore the framework

01 / Overview

Adapt Generalist VLA
To Humanoids.

Abstract

Humanoid robots hold great promise as general-purpose agents in human-centered environments, yet generalist vision-language-action foundation models are not readily applicable to humanoid whole-body loco-manipulation. High-dimensional and interdependent whole-body motion makes it difficult for conventional single-stage architectures to coordinate locomotion, waist posture, and dual-arm manipulation, while policies trained purely through offline behavior cloning can remain suboptimal during real-world deployment.

HAF, the Humanoid Adaptation Framework, combines two complementary components. HAF-VLA splits full-body action denoising into three sequential stages and preserves kinematic dependencies with stage embeddings and cross-stage KV caches. On top of the frozen generator, HAF-Steer uses flow reversal and DCT compression to train a regularized SAC policy in a compact noise subspace, enabling efficient offline-to-online refinement without updating the large VLA backbone.

Overview of HAF real-world experiments, HAF-VLA, HAF-Steer, and their performance — **Figure 1.** HAF couples hierarchical whole-body generation with spectral latent policy adaptation.

02 / Method

Two Components:
HAF-VLA &
HAF-Steer

Part A

HAF-VLA

Hierarchical Action Flow

A shared action expert progressively expands the active whole-body action space. Clean actions from earlier stages are encoded as KV caches, conditioning later stages before a single final action chunk is deployed.

Locomotion + Head

+ Waist

+ Manipulation

HAF-VLA architecture with three hierarchical action generation stages — **HAF-VLA.** Cumulative action masks and cross-stage clean-action KV caches preserve the dependency from base stabilization to fine manipulation.

Part B

HAF-Steer

Spectral Latent RL

Expert action chunks are reversed through the frozen flow field, compressed to low-frequency DCT coefficients, and used to initialize a regularized offline-to-online SAC policy.

HAF-Steer flow reversal, DCT compression, reinforcement learning, and reconstruction pipeline — **HAF-Steer.** Flow reversal constructs expert spectral targets; mixed offline-online learning adjusts the low-frequency noise modes while the flow generator remains fixed.

03 / Real-world tasks

Seven tasks.
Whole-body coordination throughout.

Each task combines navigation, posture adjustment, and physical interaction. All task clips below intentionally play without sound.

Laundry Loading

Carry + bimanual loading

Silent

Clothes Retrieval

Navigate + retrieve

Silent

Table Tidy

Reach + rearrange

Silent

Basket Transfer

Lift + transport

Silent

Toy Storage

Pick + relocate

Silent

Ball Tossing

Align + throw

Silent

Box Transfer

Squat + whole-body carry

Silent

04 / Experiments

Clear gains in both VLA & RL adaptation.

HAF-VLA Performance

HAF-VLA enable long-horizon humanoid loco–manipulation beyond existing SOTA methods

Main results on seven long-horizon humanoid loco-manipulation tasks
Task	HAF-VLA	pi0.5	GR00T N1.7	Cosmos	ACT
Laundry Loading	66.7	53.3	40.0	0.0	10.0
Clothes Retrieval	53.3	53.3	33.3	26.7	23.3
Table Tidy	80.0	70.0	40.0	16.7	23.3
Basket Transfer	63.3	50.0	43.3	33.3	16.7
Toy Storage	80.0	53.3	30.0	40.0	23.3
Ball Tossing	56.7	33.3	36.7	3.3	30.0
Box Transfer	93.3	60.0	43.3	73.3	50.0
Average	70.5	53.3	38.1	27.6	25.2

OOD robustness

HAF-VLA generalize to unseen visual and positional disturbances

With an unseen chair near the Laundry Loading path, HAF-VLA scores 40.0% versus 26.7% for pi0.5. With a 20 cm backward start shift in Clothes Retrieval, it scores 43.3% versus 36.7%.

Object distraction and position disturbance generalization experiments — Controlled visual and positional disturbances used for the HAF-VLA generalization study.

Offline to online

HAF-Steer improve real-world deployment performance through offline-to-online adaptation.

Across Toy Storage and Basket Transfer with both pi0.5 and HAF-VLA backbones, offline noise behavior cloning provides a strong initialization and offline-to-online SAC delivers further adaptation.

HAF-Steer success rates on Toy Storage and Basket Transfer in in-distribution and out-of-distribution settings — Successful trials out of 10. DSRL training was terminated in unsafe exploration settings, as noted in the paper.

HAF

Adapt Generalist VLATo Humanoids.

Two Components:HAF-VLA &HAF-Steer

HAF-VLA

Locomotion + Head

+ Waist

+ Manipulation

HAF-Steer

Seven tasks.Whole-body coordination throughout.

Clear gains in both VLA & RL adaptation.

HAF-VLA enable long-horizon humanoid loco–manipulation beyond existing SOTA methods

HAF-VLA generalize to unseen visual and positional disturbances

HAF-Steer improve real-world deployment performance through offline-to-online adaptation.

Adapt Generalist VLA
To Humanoids.

Two Components:
HAF-VLA &
HAF-Steer

Seven tasks.
Whole-body coordination throughout.