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 (arXiv 2604.19728) that unifies LLM, VLM, and VLA training in a single codebase. Unlike most open-source VLA efforts that focus only on the action-training stage and stitch together incompatible pretraining pipelines, VLA Foundry provides a shared training stack with end-to-end control, from language pretraining to action-expert fine-tuning. It 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 via an LLM-to-VLM-to-VLA pipeline, and another built on the pretrained Qwen3-VL backbone. Closed-loop policy performance is evaluated on LBM Eval, an open-data, open-source simulator. In nominal evaluation settings, the fully open-source from-scratch model matches prior closed-source work, while swapping in the Qwen3-VL backbone yields strong multi-task tabletop manipulation policies that substantially surpass baselines. Code is available at https://github.com/TRI-ML/vla_foundry, and all multi-task model weights are released on Hugging Face.

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 an LLM -> VLM -> VLA pipeline. 2. A model built on the pretrained Qwen3-VL backbone.

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

Key Results

In nominal evaluation settings, the fully open-source from-scratch model matches prior closed-source work.
Swapping in the Qwen3-VL backbone produces strong multi-task tabletop manipulation policies that substantially surpass baselines.

Resources

Code: https://github.com/TRI-ML/vla_foundry
Model weights: https://huggingface.co/collections/TRI-ML/vla-foundry
Project page (qualitative videos): https://tri-ml.github.io/vla_foundry

--- *Auto-collected on 2026-04-23*