Project · Active

Scale-CDA: democratizing AI-assisted cooperative driving automation for production cars

PI

What it is

Scale-CDA is an open-hardware / open-software toolchain that turns an ordinary production car into a working platform for generative-AI-assisted cooperative driving automation (CDA). It is built on two community-maintained foundations — the OpenDBC interface, which covers 300+ car models, and Openpilot Level-2 ADAS — so retrofitting is plug-and-play using off-the-shelf parts: an edge PC, a webcam, a CAN adapter, and optional LTE / Wi-Fi radios, for under US $1,000 total.

The stack has three layers:

  • Connectivity. A lightweight vehicle-to-everything (V2X) stack running MQTT over Wi-Fi 6 / LTE provides bidirectional messaging between vehicles and infrastructure.
  • Intelligence. An edge-deployed multimodal large language model ingests synchronized vision, CAN, and V2X streams through a Model-Context-Protocol (MCP) bridge, then emits structured JSON advisories and motion primitives. Inference stays on-board, so driving data never leaves the vehicle.
  • Actuation. A library of meta-action executors translates those high-level commands into verified Openpilot planner hooks — lane changes, gap management, emergency stops — without modifying the safety-certified core.

Field results

Testing on a 7.5 km loop measured mean round-trip latency of 5.25 ms and link speeds near 100 Mb/s, which supports Wi-Fi 6 as a viable low-cost medium for non-safety-critical CDA messaging. In multi-vehicle road tests, the full perception → reasoning → action stack held end-to-end decision latency below 60 ms.

Why it matters

CDA research has been gated by two practical barriers rather than theoretical ones. First, the hardware: cooperative-driving field trials have historically required instrumented research vehicles, which caps experiments at a handful of cars and puts large-scale deployment studies out of reach for most agencies and labs. Second, the interface: there has been no standardized way for a generative AI model to both reason about and act on an everyday vehicle without rewriting its safety-critical control path.

Scale-CDA targets both. Cheap, interoperable retrofits make fleet-scale field trials tractable, and the MCP bridge plus meta-action executors give GenAI a defined, auditable place to intervene. Releasing the bills-of-materials, connectivity APIs, and AI bridges openly gives transportation agencies and researchers a blueprint they can rebuild rather than a demo they can only read about.

Connections

Scale-CDA extends the affordable-hardware, open-source-software approach the lab developed in AI-CDA4All, and shares its instrumentation lineage with the OpenLKA data-collection work. It sits at the intersection of the lab’s AI for mobility and Vehicle technologies threads.

Project team

  • Hao Zhou (PI)
  • Shengming Yuan
  • Yuhang Wang
  • Haibin Wen

Outputs & releases

  • Paper: Zhou, Yuan, Wang, Wen. Scale-CDA: a scalable prototype to democratize AI-assisted cooperative driving automation (CDA) for production cars. arXiv:2608.04235 (Aug 2026)
  • Hardware: bill-of-materials, connectivity APIs, and GenAI bridges released as open resources — TODO Hao: link the repo under github.com/MOTIF-Lab when public