APT: Action Expert Pretraining Improves Instruction Generalization of Vision-Language-Action Policies

Hugging Face Daily Papers Papers

Summary

Researchers propose APT, a two-stage training method that pretrains action experts on vision-action pairs before integrating language conditioning, significantly improving out-of-distribution instruction generalization for Vision-Language-Action policies.

Vision-Language-Action (VLA) models that couple pretrained Vision-Language Models (VLMs) with continuous action experts have achieved strong manipulation performance, yet generalization to out-of-distribution (OOD) language instructions remains poor. A known challenge is the structural imbalance in VLA data, where language is far less diverse than visual and action content, making policies prone to visual shortcuts. While discrete-action methods mitigate this through vision-language co-training, continuous action experts lack such protection: they start from random initialization and learn entirely from imbalanced data, producing noisy gradients that corrupt the VLM and fail to exploit its language capability. We address this from a Bayesian perspective, factorizing the policy into a language-agnostic Vision-Action (VA) prior and a language-conditioned VLA likelihood, and propose APT, a two-stage training method emphasizing Action expert PreTraining. In Stage 1, the action expert is pretrained as a VA prior on vision-action pairs from a frozen VLM, bypassing the language imbalance. In Stage 2, language tokens are injected through a gated fusion mechanism that integrates VLM features while preserving the learned visuomotor prior. APT applies to mainstream VLA architectures, including the π and GR00T-style architectures. Comprehensive experiments validate that APT achieves consistent gains on unseen instructions and compositional tasks. Project Page: https://xukechun.github.io/papers/APT/
Original Article
View Cached Full Text

Cached at: 06/15/26, 12:58 PM

Paper page - APT: Action Expert Pretraining Improves Instruction Generalization of Vision-Language-Action Policies

Source: https://huggingface.co/papers/2606.12366

Abstract

Researchers address poor generalization in Vision-Language-Action models by proposing APT, a two-stage training method that pretrains action experts using vision-action pairs before integrating language conditioning to improve out-of-distribution instruction performance.

Vision-Language-Action (VLA) models that couple pretrainedVision-Language Models(VLMs) withcontinuous action expertshave achieved strong manipulation performance, yet generalization to out-of-distribution (OOD) language instructions remains poor. A known challenge is thestructural imbalancein VLA data, where language is far less diverse than visual and action content, making policies prone to visual shortcuts. While discrete-action methods mitigate this throughvision-language co-training,continuous action expertslack such protection: they start from random initialization and learn entirely from imbalanced data, producing noisy gradients that corrupt the VLM and fail to exploit its language capability. We address this from aBayesian perspective, factorizing the policy into a language-agnostic Vision-Action (VA) prior and alanguage-conditioned VLA likelihood, and propose APT, atwo-stage trainingmethod emphasizingAction expert PreTraining. In Stage 1, the action expert is pretrained as a VA prior onvision-action pairsfrom afrozen VLM, bypassing the language imbalance. In Stage 2, language tokens are injected through agated fusion mechanismthat integrates VLM features while preserving the learned visuomotor prior. APT applies to mainstream VLA architectures, including the π and GR00T-style architectures. Comprehensive experiments validate that APT achieves consistent gains on unseen instructions andcompositional tasks. Project Page: https://xukechun.github.io/papers/APT/

View arXiv pageView PDFProject pageGitHubAdd to collection

Get this paper in your agent:

hf papers read 2606\.12366

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2606.12366 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2606.12366 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2606.12366 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

Learning Agentic Policy from Action Guidance

arXiv cs.CL

The paper proposes ActGuide-RL, a method for training agentic policies in LLMs by using human action data as guidance to overcome exploration barriers in reinforcement learning without extensive supervised fine-tuning.