Risk-bounded autonomous information gathering for localization of phenomena in hazardous environments

Abstract

Exploration of new environments is often conducted in search of some phenomenon of interest. Examples include the search for extreme forms of life in the deep ocean or under the ice on Europa, or localizing resource deposits on the ocean floor. Exploration of all these environments is dangerous because of uncertainty in the environment and poorly characterized disturbances that can damage the exploration vehicle. Autonomous vehicles allows exploration in those environments where it is too dangerous or expensive to send a human-operated craft. Autonomous exploration has been well-studied from the perspective of information maximization, but information gathering has not been considered with the intention of localizing specific phenomena, nor has it been considered in environments where exploration can threaten the vehicle. This thesis addresses both challenges by introducing Risk-Bounded Adaptive Search, which maximizes the number of phenomena located while bounding the probability of mission failure by a user-defined threshold. The first innovation of this thesis is the development of a new information measure that focuses on locating instances of a specific phenomenon. Search for phenomena of interest is framed as a discrete space Markov Decision Process that is solved using forward search and receding horizon planning, with a reward function specified as the information gained about unobserved instances of the phenomenon of interest from measurements. Using this reward function, the number of phenomena located is increased compared to maximizing conventional information, as it steers the agent towards locations where phenomena are thought to exist so they are not bypassed when the belief state is high. The second innovation is a method of applying risk bounds as a function of the expected information gain of a policy over a planning horizon, in contrast to a static bound. This 'Performance-Guided Risk Bounding' system allows an MDP policy to be found that is slightly suboptimal if it has a substantially lower probability of failure, or accept more risk if the reward payoff is large. When applied to information gathering, it allows an autonomous agent to capitalize on high risk and high reward opportunities when they are seen, instead of ignoring them in an effort to conserve risk for the future, when it is ultimately less useful. Since interesting phenomena are often found in risky locations, the ability to take more risk when it is worthwhile results in more phenomena found overall. Finally, a modification to Monte Carlo Tree Search is introduced that implements Performance-Guided Risk Bounding. This allows Risk-Bounded Adaptive Search to be planned in an anytime manner. The output policy is limited to the states that are explored, but risk bounds that scale with the expected information gained over the explored states in the policy are still applied. The resulting policies are shown to converge to the results of forward search, and a few percent differences in phenomena found with an order of magnitude reduction in planning time.