Title: Towards Cloud-Assisted Autonomous Driving
Speaker: Peter Schafhalter
Advisor: Ion Stoica and Joseph E. Gonzalez
Date: Thursday, April 10, 2025
Time: 1:00pm – 2:00pm Pacific Time
Location (In-person): Soda Hall, 465H (Sky Lab)
Abstract:
Autonomous driving requires the execution of highly accurate models in real time. However, the processing power available to autonomous vehicles (AVs) is fundamentally limited by physical and economic constraints, preventing AVs from executing highly accurate, state-of-the-art models. To address these limitations, we turn to cloud computing which offers orders of magnitude more processing power at affordable prices. This talk discusses three key advancements that enable cloud-assisted autonomous driving. First, we introduce the Dynamic Deadline-Driven programming model which provides abstractions for adapting computation under dynamic runtime constraints. Next, we extend this model to implement Speculative Cloud Execution and demonstrate the feasibility of using the cloud to improve driving safety. Finally, we present TURBO, an accuracy-aware bandwidth allocation algorithm, to manage contention on the network between different AV services.