Continuous observation planning for autonomous exploration

Abstract

Many applications of autonomous robots depend on the robot being able to navigate in real world environments. In order to navigate or path plan, the robot often needs to consult a map of its surroundings. A truly autonomous robot must, therefore, be able to drive about its environment and use its sensors to build a map before performing any tasks that require this map. Algorithms that control a robot's motion for the purpose of building a map of an environment are called autonomous exploration algorithms. Because resources such as time and energy are highly constrained in many mobile robot missions, a key requirement of autonomous exploration algorithms is that they cause the robot to explore efficiently. Planning paths to candidate observation points that will lead to efficient exploration is challenging, however, because the set of candidates, and, therefore, the robot's plan, change frequently as the robot adds information to the map. The main claim of this thesis is that, in situations in which the robot discerns the large scale structure of the environment early on during its exploration, the robot can produce paths that cause it to explore efficiently by planning observations to make over a finite horizon. Planning over a finite horizon entails finding a path that visits candidates with the maximum possible total utility, subject to the constraint that the path cost is less than a given threshold value. Finding such a path corresponds to solving the Selective Traveling Salesman Problem (S-TSP) over the set of candidates.

(cont.) In this thesis, we evaluate our claim by implementing full horizon, finite horizon, and greedy approaches to planning observations, and comparing the efficiency of these approaches in both real and simulated environments. In addition, we develop a new approach for solving the S-TSP by framing it as an Optimal Constraint Satisfaction Problem (OCSP).