Hybrid Edge-Cloud Generative Pipelines for Scalable Autonomous Vehicle Fleets

Abstract

Autonomous vehicle (AV) fleets must handle vast scenario variations—from routine road scenes to rare edge cases—while balancing real-time responsiveness and computational constraints. Traditional simulation or cloud-only generative systems often face limitations in latency, bandwidth, or scalability. We propose a Hybrid Edge-Cloud Generative Pipeline (HECGP) designed to orchestrate generative scenario creation and vehicle-in-the-loop testing across AV fleets, leveraging both local compute (edge) and centralized cloud resources.
HECGP features a three-tier architecture: (1) Edge-Level Generative Agents embedded in AVs or edge nodes deploy lightweight, conditional generative models to craft immediate, context-aware scenarios (e.g., sensor noise variants, pedestrian behaviors). (2) A Cloud Aggregation and Expansion Layer collects those edge-generated seeds, enriches with high-fidelity generative models (e.g., large-scale GANs), and orchestrates large batch simulations across virtual fleets. (3) A Feedback and Deployment Engine disseminates optimized scenario parameters and updated generative models back to edge units.
Experiments show HECGP achieves up to 60% reduction in scenario generation latency relative to purely cloud-centric pipelines, while preserving high realism as evaluated by human experts. Network bandwidth usage was reduced by 40%, and fleet-wide scenario coverage increased by 30% due to dynamic edge seeding. The hybrid design further enabled scalable, privacy-preserving scenario generation, as locally produced edge seeds need not transmit raw sensor data upstream.
HECGP offers a promising paradigm for generative scenario pipelines in AV fleets—delivering flexible, low-latency scenario generation while harnessing cloud scale for depth and diversity. This architecture supports real-time adaptability, distributed learning, and fleet-scale validation, advancing AV robustness and operational readiness.