Peter Schafhalter

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2025-162

August 15, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-162.pdf

Autonomous Vehicles (AVs) have the potential to reshape transportation by eliminating human error from driving, improving traffic flow, and providing mobility to millions of people with disabilities. However, deploying AVs across a wide range of operating environments poses significant challenges due to the need to make highly accurate decisions within strict deadlines. The competing demands for high accuracy and rapid response times form a fundamental tradeoff that AV systems must navigate to ensure safety: more accurate AI models exhibit larger response times whereas less accurate models are typically faster.

In this work, we address the critical lack of techniques to reason about the tradeoff between response time and accuracy and propose several approaches that exploit this tradeoff to improve the accuracy of AV decision-making. We introduce Pylot, an open-source AV platform which provides a testbed for studying the effects of this tradeoff on end-to-end driving performance. To maximize the accuracy of decisions under the dynamically-varying deadlines imposed the environment, we propose the Dynamic Deadline-Driven (D3) execution model which centralizes deadline management. We implement D3 in ERDOS, our high-performance streaming system, which reduces collisions by 68% over prior execution models.

To address the compute constraints of AV hardware, we turn the cloud to access orders of magnitude more processing power. Speculative Cloud Execution augments AVs with cloud resources and strictly improves safety despite relying on an unreliable network. TURBO enables cloud-assisted execution of multiple services on a single AV by optimally allocating bandwidth to maximize vehicle-wide accuracy.

We believe that integrating the cloud into autonomous driving has the potential to improve safety and enable the deployment of AVs across a range of challenging driving environments. To that effect, the contributions of Pylot, D3, ERDOS, Speculative Cloud Execution, and TURBO demonstrate the advantages of integrating the cloud into autonomous driving and form a foundation for building cloud-assisted AVs.

Advisors: Ion Stoica and Joseph Gonzalez

---

BibTeX citation:

@phdthesis{Schafhalter:EECS-2025-162,
    Author= {Schafhalter, Peter},
    Title= {Towards Cloud-Assisted Autonomous Driving},
    School= {EECS Department, University of California, Berkeley},
    Year= {2025},
    Month= {Aug},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-162.html},
    Number= {UCB/EECS-2025-162},
    Abstract= {Autonomous Vehicles (AVs) have the potential to reshape transportation by eliminating human error from driving, improving traffic flow, and providing mobility to millions of people with disabilities. However, deploying AVs across a wide range of operating environments poses significant challenges due to the need to make highly accurate decisions within strict deadlines. The competing demands for high accuracy and rapid response times form a fundamental tradeoff that AV systems must navigate to ensure safety: more accurate AI models exhibit larger response times whereas less accurate models are typically faster.

In this work, we address the critical lack of techniques to reason about the tradeoff between response time and accuracy and propose several approaches that exploit this tradeoff to improve the accuracy of AV decision-making. We introduce Pylot, an open-source AV platform which provides a testbed for studying the effects of this tradeoff on end-to-end driving performance. To maximize the accuracy of decisions under the dynamically-varying deadlines imposed the environment, we propose the Dynamic Deadline-Driven (D3) execution model which centralizes deadline management. We implement D3 in ERDOS, our high-performance streaming system, which reduces collisions by 68% over prior execution models.

To address the compute constraints of AV hardware, we turn the cloud to access orders of magnitude more processing power. Speculative Cloud Execution augments AVs with cloud resources and strictly improves safety despite relying on an unreliable network. TURBO enables cloud-assisted execution of multiple services on a single AV by optimally allocating bandwidth to maximize vehicle-wide accuracy.

We believe that integrating the cloud into autonomous driving has the potential to improve safety and enable the deployment of AVs across a range of challenging driving environments. To that effect, the contributions of Pylot, D3, ERDOS, Speculative Cloud Execution, and TURBO demonstrate the advantages of integrating the cloud into autonomous driving and form a foundation for building cloud-assisted AVs.},
}

EndNote citation:

%0 Thesis
%A Schafhalter, Peter 
%T Towards Cloud-Assisted Autonomous Driving
%I EECS Department, University of California, Berkeley
%D 2025
%8 August 15
%@ UCB/EECS-2025-162
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-162.html
%F Schafhalter:EECS-2025-162