Wesley Hsieh

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2017-102

May 12, 2017

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-102.pdf

Autonomous driving is a complex task that features high variability in everyday driving conditions due to the presence of vehicles and different road conditions. Many data-centric learning models require exposure to a high number of data points collected in various driving conditions to be robust to this variability. We present the First Order Driving Simulator (FODS), an open-source lightweight driving simulator designed for data collection and benchmarking performance for autonomous driving experiments, with a focus on customizability and speed. The car model is controlled using steering, acceleration, and braking as inputs. The car features the choice between kinematic and dynamic bicycle models with slip and friction, as a first-order approximation to the dynamics of a real car. Users can customize features including the track, vehicle placement, and other initial conditions of the environment, as well as environment interface features such as the state space (images, positions and poses of cars) and action space (discrete or continuous controls, limits). We benchmark our performance against other simulators of varying degrees of complexity, and show that our simulator matches or outperforms their speeds of data collection. We also feature parallelization with Ray, a distributed execution framework aimed at making it easy to parallelize existing codebases, which allows for significant speed increases in data collection. Finally, we also perform experiments analyzing the performance of various imitation learning and reinforcement learning algorithms on our simulator environment.

Advisors: Ken Goldberg

---

BibTeX citation:

@mastersthesis{Hsieh:EECS-2017-102,
    Author= {Hsieh, Wesley},
    Title= {First Order Driving Simulator},
    School= {EECS Department, University of California, Berkeley},
    Year= {2017},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-102.html},
    Number= {UCB/EECS-2017-102},
    Abstract= {Autonomous driving is a complex task that features high variability in everyday driving conditions due to the presence of vehicles and different road conditions. Many data-centric learning models require exposure to a high number of data points collected in various driving conditions to be robust to this variability. We present the First Order Driving Simulator (FODS), an open-source lightweight driving simulator designed for data collection and benchmarking performance for autonomous driving experiments, with a focus on customizability and speed. The car model is controlled using steering, acceleration, and braking as inputs. The car features the choice between kinematic and dynamic bicycle models with slip and friction, as a first-order approximation to the dynamics of a real car. Users can customize features including the track, vehicle placement, and other initial conditions of the environment, as well as environment interface features such as the state space (images, positions and poses of cars) and action space (discrete or continuous controls, limits). We benchmark our performance against other simulators of varying degrees of complexity, and show that our simulator matches or outperforms their speeds of data collection. We also feature parallelization with Ray, a distributed execution framework aimed at making it easy to parallelize existing codebases, which allows for significant speed increases in data collection. Finally, we also perform experiments analyzing the performance of various imitation learning and reinforcement learning algorithms on our simulator environment.},
}

EndNote citation:

%0 Thesis
%A Hsieh, Wesley 
%T First Order Driving Simulator
%I EECS Department, University of California, Berkeley
%D 2017
%8 May 12
%@ UCB/EECS-2017-102
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2017/EECS-2017-102.html
%F Hsieh:EECS-2017-102