Environment Search Planning Subject to High Robot Localization Uncertainty

As robots find applications in more complex roles, ranging from search and rescue to healthcare and services, they must be robust to greater levels of localization uncertainty and uncertainty about their environments. Without consideration for such uncertainties, robots will not be able to compensate accordingly, potentially leading to mission failure or injury to bystanders. This work addresses the task of searching a 2D area while reducing localization uncertainty. Wherein, the environment provides low uncertainty pose updates from beacons with a short range, covering only part of the environment. Otherwise the robot localizes using dead reckoning, relying on wheel encoder and yaw rate information from a gyroscope. As such, outside of the regions with position updates, there will be unconstrained localization error growth over time. The work contributes a Belief Markov Decision Process formulation for solving the search problem and evaluates the performance using Partially Observable Monte Carlo Planning (POMCP). Additionally, the work contributes an approximate Markov Decision Process formulation and reduced complexity state representation. The approximate problem is evaluated using value iteration. To provide a baseline, the Google OR-Tools package is used to solve the travelling salesman problem (TSP). Results are verified by simulating a differential drive robot in the Gazebo simulation environment. POMCP results indicate planning can be tuned to prioritize constraining uncertainty at the cost of increasing path length. The MDP formulation provides consistently lower uncertainty with minimal increases in path length over the TSP solution. Both formulations show improved coverage outcomes