AutoVLA

Why we included this project

AutoVLA is the official NeurIPS 2025 implementation of a vision-language-action model that maps camera input straight to driving trajectories, so it is aimed at researchers and engineers working on end-to-end autonomous driving rather than general-purpose vision or chatbot work. The central idea is a model trained with supervised fine-tuning to switch between two thinking modes: a fast trajectory-only mode for simple scenes and a slower chain-of-thought mode for complex ones, with GRPO-based reinforcement fine-tuning deciding when the extra reasoning is worth the compute. The repository covers the full pipeline, from preprocessing nuPlan, nuScenes, and Waymo E2E data to building the action codebook, annotating chain-of-thought, and training and evaluating the model, and released checkpoints are available. Teams building or benchmarking end-to-end driving stacks can treat it as a concrete reference for combining adaptive reasoning with reinforcement fine-tuning in a single autoregressive planner. One practical caveat: the academic software license restricts use to academic and nonprofit researchers, so commercial evaluation requires contacting the lab.