CMU-Drive Benchmark for Cooperative Autonomous Driving

Existing Vision-Language-Action (VLA) models for autonomous driving primarily focus on individual agents, lacking robust support for cooperative perception, reasoning, and planning among multiple vehicles. This limitation hinders the development of truly connected autonomous vehicles (CAVs) capable of operating safely and efficiently in complex, multi-agent environments. To advance this field, researchers have introduced CMU-Drive, a novel closed-loop, end-to-end benchmark specifically designed for evaluating cooperative autonomous driving. It simulates safety-critical scenarios with background traffic, allowing for the assessment of multiple CAVs working together. Alongside this benchmark, they propose V2V-VLA (Vehicle-to-Vehicle Vision-Language-Action), a cooperative VLA model. V2V-VLA integrates cooperative driving into a single forward pass, enabling joint generation of driving actions, future waypoints, language reasoning, and communication policies between vehicles. Experiments on CMU-Drive establish the first benchmark and baseline for cooperative VLA driving, providing a foundational platform for future research in multi-agent, closed-loop, end-to-end cooperative autonomous driving. The code, benchmark, and model checkpoint will be publicly released.