VLA Foundry: A Unified Open-Source Framework for Training Vision-Language-Action Models

Forum topic · 小凯 · 2026-04-23

Summary

VLA Foundry is an open-source framework that unifies LLM, VLM, and VLA training within a single codebase, addressing a common limitation of open-source VLA efforts that only focus on the action training stage while stitching together incompatible pretraining pipelines. The framework provides a shared training stack with end-to-end control, from language pretraining to action-expert fine-tuning, and supports both from-scratch training and pretrained backbones from Hugging Face. The authors train and release two types of models: one trained fully from scratch through an LLM-to-VLM-to-VLA pipeline, and another built on the pretrained Qwen3-VL backbone. Both models are evaluated for closed-loop policy performance on LBM Eval, an open-data, open-source simulator. Under nominal evaluation settings, the fully open from-scratch model matches prior closed-source work, while the Qwen3-VL backbone variant yields a strong multi-task tabletop manipulation policy that substantially outperforms baselines. Code, model weights, and project videos are publicly released.

Paper Overview

Field: Computer Vision (CV)
Authors: Jean Mercat, Sedrick Keh, Kushal Arora, Isabella Huang, Paarth Shah, Haruki Nishimura, Shun Iwase, Katherine Liu
Published: 2026-04-21
arXiv: 2604.19728

Abstract

We present VLA Foundry, an open-source framework that unifies LLM, VLM, and VLA training in a single codebase. Most open-source VLA efforts specialize on the action training stage, often stitching together incompatible pretraining pipelines. VLA Foundry instead provides a shared training stack with end-to-end control, from language pretraining to action-expert fine-tuning.

VLA Foundry supports both from-scratch training and pretrained backbones from Hugging Face. To demonstrate the utility of the framework, the authors train and release two types of models:

1. A model trained fully from scratch through the LLM → VLM → VLA pipeline 2. A model built on the pretrained Qwen3-VL backbone

Both models are evaluated for closed-loop policy performance on LBM Eval, an open-data, open-source simulator. The authors also contribute usability improvements to the simulator and STEP analysis tools for public use.

Results

Under nominal evaluation settings, the fully open, from-scratch model matches prior closed-source work
Swapping in the Qwen3-VL backbone produces a strong multi-task tabletop manipulation policy that substantially outperforms baselines