Abstract
Autonomous mobile robots (AMRs) are increasingly being deployed in real-world applications such as exploration, search-and-rescue, and inspection. Yet despite significant advances in onboard computation and perception, AMRs continue to struggle with safe, reliable navigation in unknown, cluttered environments. Collisions often arise due to occlusions, environmental uncertainty, and sensor noise. These challenges are exacerbated by traditional, decoupled navigation stacks, where a low-level controller blindly tracks high-level trajectories, even when they are unsafe. This brittleness in the motion planning pipeline becomes increasingly problematic for high-speed, time-critical tasks and remains a key barrier to widespread, real-world deployment.
This dissertation focuses on the design of a robust, agile, and safe navigation stack for AMRs. In particular, this work goes through each layer of the navigation stack: waypoint generation, motion planning, and control, to develop interventions that improve their robustness without sacrificing safety or agility. First, an exploration waypoint generation approach is developed which leverages occlusion-aware navigation, where carefully selected occluded regions in the robot's onboard sensors are used to dictate unknown regions to explore. While exploring, a secondary algorithm is used online to generate a refined shortest path tour through the environment for future exploration missions. Second, failures to generate updated trajectories in the motion planning layer are cast as a Gaussian Process (GP)-based risk-aware problem which estimates the future probability of planner failure along a fixed time horizon. When the risk of failure becomes too large, the robot halts, finds a region of high planning success probability, and navigates there to continue normal operation. Lastly, at the control layer we bridge the gap between theory and application for the increasingly popular Control Barrier Function (CBF)-based safety filters. While multiple low-level controllers are investigated in this section, we focus primarily on the Model Predictive Controller (MPC), and develop a strategy for tuning the safety filter parameters in a manner that allows for agile motion without ever compromising safety. The theoretical and practical implications of these approaches are discussed and validated through extensive simulations and experiments, both in lab-controlled settings and outside the lab in real-world environments.
