Difference between revisions of "SURF 2020: Rules of the Road--a top-down approach for guaranteeing correct behavior for self-driving cars"

From MurrayWiki
Jump to: navigation, search
(Created page with "The current prevailing approach for certifying the correctness of the behavior of autonomous vehicles is to simulate or test drive the vehicle for millions of miles. This bott...")
 
(No difference)

Latest revision as of 19:30, 3 December 2019

The current prevailing approach for certifying the correctness of the behavior of autonomous vehicles is to simulate or test drive the vehicle for millions of miles. This bottom-up approach is not only inefficient and impractical, but it also means manufacturers cannot provide formal guarantees for critical behavioral specifications like safety.

We have proposed a top-down approach for designing the behavioral specifications of an autonomous vehicle. In particular, we have designed an assume-guarantee paradigm under which each vehicle is assumed to act according to some behavioral profile. Under the assumptions that all cars are operating according to these specific profiles, the car can guarantee it will behave according to its own behavioral profile. Each behavioral profile has specific mathematical properties that function to order the importance of all desired behavioral specifications in an autonomous vehicle.

Slide.png

With each car acting according to this type of behavioral profile, we can begin to guarantee certain behavioral specifications of each car. The objectives of this project would be to:
1. Prove and/or simulate the collective behavior that emerges when each car is acting according to behavioral specifications in all different road network configurations and initial conditions.
2. Reduce the algorithmic complexity for resolving games with multiple agents so that it scales well.
3. Show the differences (advantages and disadvantages) between thinking of all cars collectively (assume-guarantee profiles) vs. as individual entities.

Note, we have already constructed a simulation environment to run and test algorithms, so familiarity with the following topics would be of great help:
- Python
- Game theory
- Dynamic Programming and Optimization

References
[1] Phan, T. M., Cai, K. X. and Murray, R. M. Towards Assume-Guarantee Profiles for Autonomous Vehicles, Control and Decision Conference 2019. https://arxiv.org/abs/1909.04850
[2] Censi, A. Slutsky, K. Wongpiromsarn, T., Yershov, D., Pendleton, S., Fu, J. and Frazzoli, E. Liability, Ethics, and Culture-Aware Behavior Specification using Rulebooks, International Conference on Robotics and Automation 2019. https://arxiv.org/pdf/1902.09355.pdf