Towards Reliable Navigation in Modern Transportation Systems

Abstract: Navigation in modern transportation systems involves interacting with large numbers of vehicles, both human-operated and self-driving, which raises serious issues of safety and efficiency. First, autonomous vehicles remain unable to consistently operate safely in real-life traffic. Key reasons include the difficulty of performing fast and accurate state estimation and mapping, as well as the absence of feedback loops between behavior prediction and motion planning in many autonomy stacks. Second, high traffic loads induce societal-scale externalities, such as excessive commute times and pollution. Unfortunately, conventional mechanisms for congestion management are often computationally intractable and ignore changes in travelers’ route selections over time.

In this talk, I will present progress towards addressing these issues, focusing in particular on: (I) A unified optimization-based state estimation and mapping framework for robotics tasks; (II) Game-theoretic motion planners that capture feedback loops between behavior prediction and path planning in common traffic scenarios; and (III) A dynamic tolling scheme to manage congestion on traffic networks of arbitrary structure and scale. I will conclude by describing promising avenues of future work.

Towards Reliable Navigation in Modern Transportation Systems

Abstract: Navigation in modern transportation systems involves interacting with large numbers of vehicles, both human-operated and self-driving, which raises serious issues of safety and efficiency. First, autonomous vehicles remain unable to consistently operate safely in real-life traffic. Key reasons include the difficulty of performing fast and accurate state estimation and mapping, as well as the absence of feedback loops between behavior prediction and motion planning in many autonomy stacks. Second, high traffic loads induce societal-scale externalities, such as excessive commute times and pollution. Unfortunately, conventional mechanisms for congestion management are often computationally intractable and ignore changes in travelers’ route selections over time.

In this talk, I will present progress towards addressing these issues, focusing in particular on: (I) A unified optimization-based state estimation and mapping framework for robotics tasks; (II) Game-theoretic motion planners that capture feedback loops between behavior prediction and path planning in common traffic scenarios; and (III) A dynamic tolling scheme to manage congestion on traffic networks of arbitrary structure and scale. I will conclude by describing promising avenues of future work.