arXiv:2608.07621 Releases CMU-Drive Benchmark and V2V-VLA Model for Multi-Agent CAV Driving

The paper defines CMU-Drive as a closed-loop evaluation suite focused on safety-critical scenarios involving multiple CAVs and background traffic. V2V-VLA extends single-agent VLA architectures by embedding cooperative perception and joint policy generation directly into the model. Experiments establish baseline performance for end-to-end cooperative driving without modular handoffs between perception, planning, and communication modules.

Related work on single-vehicle VLAs such as DriveVLM and LanguageMPC shows strong isolated-agent results yet lacks explicit multi-agent coordination mechanisms. CMU-Drive fills this gap by requiring agents to share partial observations and align plans under occlusion and interaction constraints. The benchmark records joint metrics including collision rate, throughput, and communication bandwidth, exposing trade-offs absent from prior single-agent suites.

Operational impact centers on reduced latency for cooperative maneuvers in dense traffic, where V2V-VLA replaces sequential V2X message passing with a unified forward pass. Public release of code, benchmark scenarios, and model weights will allow direct replication and extension by other groups working on multi-agent systems.

Future deployments will test whether single-pass cooperative VLAs maintain performance when scaled to larger vehicle fleets and real-world sensor noise distributions.