StarVLA: VLA Model Codebase

Technical Analysis Open-source codebases have become very popular and useful in the robotics community, as they simplify, unify, and speed up model building and trainig. Here comes "StarVLA: A Lego-like Codebase for Vision-Language-Action Model Developing", a modular and flexible codebase for developing Vision-Language Model (VLM) to Vision-Language-Action (VLA) models. Motivation of the Work The primary motivation is to create a modular and flexible infrastructure that makes VLA model development "Lego-like". This structure is intended to allow for plug-and-play design, rapid prototyping, and independent debugging of different components. Key contributions include: A highly modular codebase: StarVLA follows principles of high cohesion and low coupling, separating core functional components (model, data, trainer, evaluation, etc.) in an intuitive, top-down manner. Rapid Development: The design has been beta-tested internally, showing that a new VLA framework can be stood up in under half a day, and a new user can build their first custom framework within a single day. Framework Support: The system includes ready-to-use implementations of several representative VLA frameworks, often using Qwen VLMs as the foundation (e.g., Qwen-FAST, Qwen-OFT, Qwen-PI, and Qwen-GR00T). Open Access: StarVLA is released under the permissive MIT License, along with a series of pre-trained models and checkpoints (Model Zoo) to help others reproduce results and build new systems. Contribution and…