Scalable Planning Solutions for Autonomous Vehicles
This research presents MODIA—a scalable collection of POMDPs active simultaneously, one for each aspect of the situation—successfully implemented on Nissan's fully operational autonomous vehicle prototype for urban intersection decision-making with vehicles and pedestrians.
Multi-Objective Sequential Optimization
Real-world solutions often require optimizing using multiple objectives. This work proposes a set of new multi-objective MDP-based models that sequentially solve each objective, allowing for slack--allowable deviation from optimal--to be applied at each step. This enables us, for example, to devise a solution to complex multi-objective hybrid vehicle engine activation in which we minimize energy consumed, noise produced, and the number of engine mode transitions to reduce emissions and more smoothly integrate these vehicles into society.
High Performance Planning Algorithms
Given the complexity of SSP and POMDPs, this line of research focuses on designing scalable and efficient planning algorithms for these models. For POMDPs, leveraging novel theoretically-grounded controller representations shows dramatic performance improvement with reduced memory usage. For SSPs, fast labeling and sampling techniques have been employed for both single- and multi-objective scenarios. These algorithms are shown to outperform other state-of-the-art approaches and are demonstrated on real robots.
“Log-Space Harmonic Function Path Planning” IROS 2016
“Fast SSP Solvers Using Short-Sighted Labeling” AAAI 2017
“Scalable POMDP Decision-Making Using Circulant Controllers” ICRA 2021
Safety in Semi-Autonomous Systems
Safety is a major aspect of any deployed AI system. This work focuses on the shared control of a system, such as in current- and next-generation autonomous vehicles. Hierarchical approaches are used to strive towards a safe transfer of control. Competence awareness of the robot enables it to proactively request for help, obviating unsafe situations, with the ability to learn to improve its model over time. Applications include semi-autonomous vehicle route planning and campus robot navigation, with theoretical guarantees such as a notion of provable safety within the model.
