Learning to Race Full-Scale Autonomous Racecars

Eric Berndt

EECS Department
University of California, Berkeley
Technical Report No. UCB/EECS-2025-120
May 16, 2025

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-120.pdf

We present a causal transformer, trained via behavior cloning to emulate a model–predictive controller, and demonstrate its zero-shot deployment on a full-scale racecar capable of driving up to 200 mp/h (320 km/h). First, we generate realistic circuits in simulation from online GPS maps, obtaining noisily-measured center-lines and track boundaries that reflect the variability of real venues. Second, we compute time-optimal race lines on every circuit with a spline-based minimum-curvature optimizer. Third, an MPC tracks each race line in simulation while action noise broadens the state distribution; the resulting state–action corpus is tokenized and a transformer is trained to predict future steering angles and drive forces. In simulation, the learned policy closely matches the expert up to 120 mp/h. The model is then deployed zero-shot on the real Las Vegas Road Course, where the network drives the autonomous racecar at 40 mp/h without crashes, with the only degradation being latency-induced oscillations that multi-step prediction alleviates. To our knowledge, this is the first demonstration of transformer-based behavior cloning for a full-scale autonomous racecar and provides the first step towards achieving a superhuman level racer.

Advisor: S. Shankar Sastry

"; ?>

BibTeX citation:

@mastersthesis{Berndt:EECS-2025-120,
    Author = {Berndt, Eric},
    Title = {Learning to Race Full-Scale Autonomous Racecars},
    School = {EECS Department, University of California, Berkeley},
    Year = {2025},
    Month = {May},
    URL = {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-120.html},
    Number = {UCB/EECS-2025-120},
    Abstract = {We present a causal transformer, trained via behavior cloning to emulate a model–predictive controller, and demonstrate its zero-shot deployment on a full-scale racecar capable of driving up to 200 mp/h (320 km/h). First, we generate realistic circuits in simulation from online
GPS maps, obtaining noisily-measured center-lines and track boundaries that reflect the variability of real venues. Second, we compute time-optimal race lines on every circuit
with a spline-based minimum-curvature optimizer. Third, an MPC tracks each race line in simulation while action noise broadens the state distribution; the resulting state–action
corpus is tokenized and a transformer is trained to predict future steering angles and drive forces. In simulation, the learned policy closely matches the expert up to 120 mp/h. The model is then deployed zero-shot on the real Las Vegas Road Course, where the network drives the autonomous racecar at 40 mp/h without crashes, with the only degradation being
latency-induced oscillations that multi-step prediction alleviates. To our knowledge, this is the first demonstration of transformer-based behavior cloning for a full-scale autonomous racecar and provides the first step towards achieving a superhuman level racer.}
}

EndNote citation:

%0 Thesis
%A Berndt, Eric
%T Learning to Race Full-Scale Autonomous Racecars
%I EECS Department, University of California, Berkeley
%D 2025
%8 May 16
%@ UCB/EECS-2025-120
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2025/EECS-2025-120.html
%F Berndt:EECS-2025-120