Title: CADET: A Modular Framework for Assessing Distributed Cooperative Autonomy in Connected Autonomous Vehicles

Abstract:

While deep learning has become a cornerstone of autonomous vehicle (AV) development, its implementation has historically adhered to a monolithic architecture. In this traditional setup, perception, planning, and control functions are consolidated onto a single onboard computer. This approach fails to account for the shifting landscape of cooperative autonomy, a paradigm where vehicles leverage vehicle-to-everything (V2X) connectivity to interact with roadside units (RSUs), edge servers, and cloud-based intelligence. Although cooperative perception and control mechanisms enhance both safety and operational efficiency, they introduce complex systems-level hurdles. Issues such as network latency, heterogeneous computing resources, and contention among multiple tenants significantly impact real-time decision-making processes. These difficulties are exacerbated by the growing dependence on large foundation models, which are too massive to run locally and thus require cloud deployment.

To address these issues, we introduce CADET (Cooperative Autonomy through Distributed Experimentation Toolkit), a modular platform designed for the systematic and reproducible assessment of distributed cooperative autonomy systems under realistic conditions. CADET separates the AV stack into distinct, composable modules, allowing for flexible deployment across vehicles, infrastructure, and edge or cloud tiers. The framework incorporates state-of-the-art models and utilizes trace-driven emulation for both network and workload simulation. Furthermore, it offers synchronized instrumentation at the model, system, and task levels.

Our experiments involving vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) communications demonstrate that deployment strategies are critical determinants of safety. Specifically, V2V intent packets proved superior to cloud-based perception methods, while RSU-assisted perception maintained safety standards until it was overwhelmed by concurrent requests. While primarily intended for AV pipelines, CADET also facilitates dataset-driven experimentation. This feature allows systems and machine learning researchers to benchmark distributed inference workloads without needing a full vehicle simulation. The CADET project is open source, with code and a demo available at https://nesl.github.io/cadet-web.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC