Robot systems have become prevalent and transformative in many areas, such as environment surveillance and reconnaissance, search and rescue, industrial manufacturing, and transportation. In these applications, it is critical for robots to navigate autonomously and reliably in the environment in order to execute their tasks. This requires efficient maintenance of an environment model, offering minimal storage footprint and fast inference time, and an accurate robot dynamics model, enabling stable and robust control policies in novel operating conditions. This dissertation proposes a novel autonomous navigation approach that utilizes machine learning techniques to develop sparse probabilistic occupancy maps of the environment and learn robot dynamics efficiently from data by preserving prior knowledge in the dynamics model.
The first part of the dissertation develops a compact machine learning model, trained online from streaming sensory data, to represent the occupancy of the environment. While common occupancy maps might have high storage requirements for large environments, we propose a novel approach that models the obstacle boundary as the decision boundary of a machine learning classifier, and thus, scales with the complexity of the boundary instead of the environment size. We develop online training algorithms of kernel perceptron and relevance vector machine classifiers to incrementally build sparse binary and probabilistic occupancy maps, respectively, from local observations.
The second part of the dissertation proposes a machine learning model for learning accurate robot dynamics from state-control trajectories. While hand-designed models might over-simplify the dynamical system, black-box models recently have become increasingly popular but require a large amount of data for training. We develop a data-efficient hybrid approach by encoding prior knowledge such as universal laws of physics and the kinematic structure of the state manifold in the dynamics model. The encoded prior knowledge is guaranteed by design instead of being inferred from data. In novel operating conditions, this approach is extended to learn a disturbance model to handle dynamics changes.
The dissertation finally develops efficient collision checking algorithms for motion planning with the learned sparse map representations and trajectory-tracking control policies based on the learned robot dynamics and disturbance models, offering a fast, reliable, and long-term solution for autonomous navigation. The autonomous navigation approach is verified extensively with datasets, simulated and real robot experiments.
