VLA Foundry Framework Unifies Vision-Language-Action Model Training with Open-Source Release

VLA Foundry introduces an open-source framework that consolidates training for large language models, vision-language models, and vision-language-action models within a single codebase. Unlike specialized approaches that focus solely on action training stages, this framework offers a comprehensive training stack with end-to-end control from initial language pretraining to specialized action-expert fine-tuning. The system supports both training models from scratch and utilizing pretrained backbones available through Hugging Face. To demonstrate its capabilities, the framework has been used to train and release two distinct model types: one trained entirely from scratch using an LLM→VLM→VLA pipeline, and another constructed on the pretrained Qwen3-VL backbone. These models underwent evaluation for closed-loop policy performance using LBM Eval, an open-data, open-source simulator. The project also contributed usability enhancements to both the simulator and STEP analysis tools to facilitate broader public adoption. The framework addresses the common problem of incompatible pretraining pipelines in existing open-source VLA efforts by providing unified infrastructure.