Long-term monitoring for robots with probabilistic models

Thesis

As robots become increasingly capable and robust, there is a growing scope for long-term monitoring missions in which robots operate autonomously without human intervention. Long-term robot missions hold immense potential across a range of scientific and industrial applications, from environmental monitoring to precision agriculture. However, mission planning in this setting is challenging, requiring robots to reason about complex and uncertain environment models over long horizons. Many existing works on planning for long-term robot missions use simplistic models of robots and their environments, overlooking the complexity and uncertainty inherent to real-world robotics in unstructured environments.

This thesis addresses planning for long-term robot monitoring of unknown, continuous spatiotemporal environment processes. We use Gaussian processes to capture the dynamics of these environment processes using sparse observational data, adopting an online approach that assumes minimal prior knowledge. We develop planning algorithms that reason about these models in belief space over long horizons to produce effective reward-gathering policies. Our algorithms address practical challenges such as resource constraints and uncertainty over action durations, which are fundamental to long-term robot missions in the real world. We also present an algorithm for safe exploration and mapping in the presence of an unknown environmental hazard, enabling a robot to autonomously and safely map an unknown environment in which no prior map is available. Our empirical results demonstrate that our planning solutions are able to generate effective policies in realistic and challenging long-term robot monitoring scenarios.

Authors: Stephens, R

• DOI: 10.5287/ora-mvyvmoa6p
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: robotics; planning under uncertainty; long-term monitoring; Markov decision processes; Gaussian processes; Monte Carlo tree search
• Subjects: Robotics