Adaptive robotic search and sampling of sparse natural phenomena

Abstract

Autonomous robots are increasingly being used in the field of scientific exploration and data acquisition. Intelligent autonomous robots, capable of online adaptive planning, are seeing wide use in underwater field mapping and agricultural monitoring. The majority of these approaches produce maps of easily observable and widely dispersed phenomena such a temperature, salinity or tree coverage. However underwater and planetary science can often involve phenomena that are ‘expensive’ to observe, discrete, and sparsely distributed. For example, coral disease can only be visually detected by an underwater robot when hovering close to the reef, due to light attenuation underwater, putting the robot at risk of collision with obstacles or organisms. Similarly, subsurface water on Mars can only be detected from a landed system on the surface, due to the short range of the detectors. When the operating conditions are resource-constrained, such as a limited battery life, expensive sensing actions can consume the resource budget, limiting the range of area that can be explored. The tension between needing to act intelligently to find and measure sparse phenomena, and needing to operate within resource constraints, leads to challenges for the robot’s autonomous decision making process in choosing what to sense, where, and when. This thesis aims to address this challenge by combining semantic ‘substrates’ in the environment with hierarchical probabilistic modelling which maps substrate distributions to the underlying phenomena of interest. By using substrates that are detectable over a wide field of view, and correlated with sparser and harder to find phenomena, a robot can be guided to regions known to be associated with the phenomena of interest. This problem can be formulated as a partially-observable Markov decision process (POMDP) referred to as the Discrete Search and Sample problem. This thesis proposes two algorithmic contributions to the field of adaptive path planning to address two scenarios within this framework. In the first scenario, we assume the robot has prior knowledge about the expected density of discrete targets in the various substrates, however is operating without prior knowledge of substrate distributions. We develop a novel multi-altitude planning method, the Sparse Adaptive Search and Sample (SASS) for seeking out targets by mixing low-altitude observations of discrete targets with high-altitude observations of the surrounding substrates. By using the prior information about the distribution of targets across substrate types in combination with belief modelling over these substrates in the environment, high-altitude observations provide information that allows SASS to quickly guide the robot to areas with high target densities. In our second scenario, the a priori assumption of substrate-target correlation models is relaxed and the robot is now operating without strong prior knowledge of target density, or the relationship between target and substrate. Drawing inspiration from the Species Distribution Modelling community, an hierarchical probabilistic model is developed using the Integrated Nested Laplace Approximation framework, that enables online inference about expected target hotspots using predicted substrate distributions. Model parameters are learned online to build a prediction over the discrete targets, and the model is integrated into an anytime online planner to enable adaptive path planning. Both algorithms are extensively evaluated with both synthetic and real-world datasets. Additionally, through the course of addressing these two scenarios, two novel generative species-substrate model were developed that enable rapid simulation of synthetic worlds, with properties derived from real-world data. The development of these simulators allow the testing of path planners that aim to exploit natural correlations in spatial distributions that occur in the real world.