WiP: Simulation and Evaluation of Collaborative Cloud-Assisted Multi-Vehicle Autonomous Driving

Abstract

Autonomous vehicles (AV) rely on compute intensive perception pipelines like detection, segmentation and prediction, that often exceed the capability of embedded vehicle hardware. Cloud-assisted AVs alleviate this by offloading computation to remote GPU servers, but as fleets grow, many vehicles contend for the same cloud and network resources. The behavior of such systems at scale remains poorly understood. In this paper, we extend CARLA and Pylot from a single-vehicle platform to a multi-vehicle simulation testbed, and present a collaborative cloud-assisted driving framework built around a dual-path perception architecture. Each vehicle maintains a complete local perception stack as a guaranteed fallback and opportunistically offloads to a cloud server running a higher-accuracy pipeline. The cloud result replaces the local result when it arrives within the planning deadline. To curb redundant perception across the fleet, a cloud-based spatiotemporal memory lets vehicles share perception outputs and thereby avoid re-running the full pipeline for obstacles already characterized by a preceding vehicle. Our preliminary experiments show that cloud inference latency stays bounded under shared execution, revealing that the cloud is not the binding bottleneck at small fleet sizes.