Prospection for Mobile Robots in Unknown Environments

Publication Type:
Thesis
Issue Date:
2023
Full metadata record
The problem of robot autonomy has been traditionally posed as that of improving adaptivity to support wider ranges of environments. This thesis argues that adaptivity alone is insufficient, because the resulting behaviour is inherently limited to mere reactions to external stimuli. Therefore, robotic prospection is proposed as a new framework for autonomy that extends to proactive behaviours. Prospection is a concept from cognitive psychology that refers to the generation and evaluation of possible future scenarios and outcomes. Robotic prospection is first mathematically formulated from a Bayesian perspective, yielding subproblems of prospective perception and planning. The prospective perception problem asks to design predictive priors that expedite Bayesian environmental perception. The prospective planning problem aims to design a strategy that maximises a task reward by, for example, balancing exploration and exploitation. To address these problems, a number of algorithmic tools are developed including mutual information upper confidence bound (MI-UCB) and random signal temporal logic (RSTL) for prospective planning, and advection-diffusion Gaussian processes (GPs), incompressible GPs, and log GP implicit surface (log-GPIS) for prospective perception. It is hoped that the formulations and solutions presented in this thesis serve as a methodical recipe for robot autonomy in the future.
Please use this identifier to cite or link to this item: