Robot Learning and Planning with a Probabilistic Perspective

Abstract

Robots excel at simplifying intricate and repetitive tasks in our daily routines. However, a more deliberate approach is essential for critical processes requiring timely and thoughtful decision-making. This thesis explores robot planning problems within high-dimensional operating domains and is characterised by combinatorial complexity. The methods presented here are firmly rooted in learning-based approaches, offering a structured framework to effectively address these intricate challenges through probabilistic modelling. At the heart of this research lies the pivotal concept of informed and learning-based planning, which equips robots with the unique ability to leverage crucial information regarding both their own characteristics and the intricate dynamics of their surrounding operational environment. By adeptly harnessing this wealth of information, robots gain the power to navigate through intricate and demanding scenarios with exceptional efficiency and effectiveness. The adoption of informed reasoning leads to a substantial enhancement of their problem-solving capabilities when confronted with the intricacies and challenges posed by complex and ever-changing environments. Informed planning, underpinned by the principles of learning and reasoning, emerges as the linchpin, elevating the potential of robotic systems to adapt, respond, and excel in addressing multifaceted real-world challenges. One significant contribution of this thesis is introducing a methodological class to address the intricacies of narrow passage motion planning. This approach employs local sampling-based planning to explore feasible trajectories within the configuration space. Local planning leverages the inherent structure of the search space via Markov chains and employs Bayesian principles to model the proposal distribution. Experimental validation was conducted in both simulated and physical environments, complemented by theoretical proofs of completeness and optimality for the methods. A second notable contribution addresses constructing learning-based probability distributions to enhance future planning capabilities. Two distinct methodologies are presented. The first method directly learns a conditional sampling distribution from historical planning data, biasing the planner towards relevant regions. The second approach constructs a diffeomorphic distribution from occupancy maps obtained from the surrounding environment, facilitating application in novel environments without prior data knowledge. Both methods were applied to optimise the distribution of various planners for manipulators. Finally, this thesis introduces another method contributing to robot learning and planning in a broader multi-trees motion planning context. Particularly challenging are regions of narrow passages needing more explicit representation, making local planning deployment non-trivial decisions. This thesis addresses this with a learning-based approach that formulates a surrogate function, enabling adaptive local tree proposals in regions where local exploitation is most advantageous. As a result, the method adeptly utilises rapid tree-based planning in open regions while deploying local planning in areas that pose connectivity challenges. Theoretical asymptotic optimality guarantees are provided, with experimental results verifying practical viability.