The effectiveness of robot autonomy is governed by the ability to make decisions based on online sensor measurements and a prior belief of the environment. Uncertainty in the environment introduces challenges to robotic decision making. This thesis address two key robot decision making problems: exploration and navigation. The robotic exploration problem requires a robot to maximally observe an unknown environment. A key challenge of robotic exploration is dealing with prior map uncertainty that arises due to partial knowledge of the world at the beginning of the mission. Since the map is unknown, it is difficult to determine which action gains the most information. The robot navigation problem is to determine a sequence of states in a given environment to move the robot from a start to goal state while avoiding collisions. Sensors mounted on real robots have measurement noise which leads to uncertainty in the mapping of obstacles. This uncertainty in obstacle location makes it challenging to find collision free paths.
Real-world environments consist of interesting topological structures such as rooms, corridors, connections, intersections, and loops. These topological structures can be exploited to address uncertainty in the environment. Computational topology methods can be utilized to robustly extract environment topology. In this thesis, we resolve robot decision-making challenges due to environment uncertainty by leveraging topological methods that are robust to noise.
We first address prior map uncertainty by predicting unknown regions of subterranean environments. Our method involves a convolutional neural network and a novel loss that combines image inpainting features and topological features via persistent homology. Simulation results on four real mines and a dataset of procedurally-generated networks show our algorithm can exploit subtle topological structural cues in subterranean environments for efficient exploration.
We then address sensor measurement uncertainty by constructing a topologically accurate roadmap from a probabilistic occupancy map.We propose an unsupervised topological learning technique with biased sampling in topologically important regions. Simulation results show a robot can search a collision-free and cost-effective path over the topological roadmap in low runtime.
The proposed topological learning is a significant contribution to the two key problems of robot decision making. Our experimental results show a robot can successfully learn to exploit topological structures for online exploration of subterranean environments.Further we show that through topological biased sampling, a robot can construct a topologically accurate roadmap for robot navigation tasks.