keivalya/mini-vla

Implements diffusion-based action generation with separate encoders for vision (images), language (text instructions), and robot state, fused via an MLP before a diffusion policy head—all contained in ~150 lines of core model code. Designed for Meta-World environments with a complete pipeline: expert data collection, training on trajectory datasets, and inference with free-form text instructions. Prioritizes educational clarity and rapid prototyping over production optimization, making it suitable for learning diffusion policies and VLA architecture without heavy framework dependencies.

204 stars.