Towards Risk-Aware Planning in Uncertain Environments

Abstract

Robotic manipulators have the potential to enhance modern healthcare by performing diagnostic procedures in point-of-care settings, assisting surgeons in operating rooms, and accelerating the discovery of novel therapeutics in research laboratories. However, researchers must address several key challenges to ensure robots can accomplish these important tasks. First, robots should be autonomous. This means robots should be capable of sensing their surroundings, gathering and learning from new information, and making their own decisions about how to accomplish specific tasks. Second, robots should be safe and only perform actions that are guaranteed to not damage objects in the environment, nearby humans, or even the robot itself. Third, robots should be robust to uncertainty. For example, a research robot should be able to clear instruments from a workbench without knowing the exact mass or friction coefficient of each instrument. Finally, robots should operate in real-time. This ensures that robots can quickly adapt their behavior to task or environment changes. The goal of this thesis is to address the challenges discussed above by developing a novel motion planning framework that integrates perception, reachability analysis, and control algorithms to generate safe trajectories in a receding horizon fashion. The proposed framework is the result of the following major contributions: (1) the development of a core trajectory planning framework based on reachability analysis; (2) the development a novel sphere-based safety representation that facilitates the integration of perception models into the planning framework; and, (3) the application of the trajectory planner as a novel differentiable neural network layer. My approach is distinctive in three ways. First, it ensures that both safety and dynamics constraints are robust to uncertainty and satisfied in continuous-time rather than discrete-time. Second, it studies how to integrate artificial intelligence, trajectory optimization, and control into a coherent framework for robot learning with theoretical guarantees. Third, by leveraging accelerated computer hardware, the proposed framework is computationally tractable and capable of being implemented in real-time on real robotic systems. In this thesis, we demonstrate this framework’s effectiveness by solving a variety of challenging motion planning and manipulation tasks in simulation and on real hardware.