Least-Violating Motion Planning for Traffic-Compliant Autonomous Driving
Time: Wed 2022-06-01 15.00
Location: F3, Lindstedtsvägen 26 & 28, Stockholm
Video link: https://kth-se.zoom.us/j/67837765464
Language: English
Subject area: Computer Science
Doctoral student: Jesper Karlsson , Robotik, perception och lärande, RPL
Opponent: Assistant Professor Tichakorn Wongpiromsarn, Iowa State University, Ames, (IA), USA
Supervisor: Jana Tumova, Robotik, perception och lärande, RPL
QC 20220516
Abstract
Over the last decade, autonomous vehicles has received an increasing amount of interest from industries and research institutes. For autonomous vehicles to properly function alongside human drivers, safety guarantees are a must. Safety in traffic is more than just avoiding collisions with other drivers, it is also necessary to seamlessly act and interact in traffic.
Traffic is an environment rife with rules, both straightforward road rules, e.g. “stay in your lane”, and more subtle road rules, e.g. “give way to emergency vehicles”. Given the safety-critical nature of the environments in which an autonomous system needs to act, it is essential that the specification language chosen to encode its behaviour is able to express the full range of possible rules, both straightforward and subtle. Linear Temporal Logic (LTL) is a popular specification language used in motion planning. While LTL is suitable to express many basic rules, more elaborate rules need to incorporate continuous measures of satisfaction. For instance, it is possible to formalize ``maintain the speed limit'' in LTL. However, there is a big difference between violating the speed limit by \SI{2}{km/h} and \SI{30}{km/h}, this difference can not be quantified by LTL. Such measures are offered by Signal Temporal Logic (STL). In our work, we have used both LTL and STL to encode complex road rules including allowable distances to obstacles, as well as more complex road rules for various situations.
The aim of our work has been to formalize and verify safety guarantees for motion planning in autonomous vehicles. This thesis’ contribution encompasses three main venues of research in this area. First, current methods employed in formal synthesis for motion planning are too computationally expensive to reliably provide motion plans in real-time. To this end, we propose solutions to two different problems, scalability and guided sampling for sampling-based motion planners (Papers A and D). Second, we deal with the problem of encoding road rules for motion planning applications. We propose a new spatial-temporal quantitative semantic for STL, that allows the user to calibrate preference for efficiency (duration of mission) against perceived safety (violation of specification)(Paper B). We later show how STL can be used to encode traffic behaviours (Paper E). Third, we investigate the problem of least-violating motion planning in mixed-traffic scenarios (Papers C and E). Here we consider two different viewpoints, humans as dynamic obstacles to avoid (Paper C) and humans as participants in traffic (Paper E). We demonstrate how least-violating motion planning combined with STL, can be utilized to encode road rules in such a way as to produce different forms of driving styles that are perceivable by human users.