Robot learning with strong priors

Abstract

Embedding learning ability in robotic systems is one of the long sought-after objectives of artificial intelligence research. Despite the recent advancements in hardware, large-scale machine learning algorithms and theoretical understanding of deep learning, it is still quite unrealistic to deploy an end-to-end learning agent in the wild, attempting to learn everything from scratch. Instead, we identify the importance of imposing strong prior knowledge on capable robotic systems and perform robot learning with strong priors. In this thesis, we exemplify the value of imposing strong priors in robot learning (or machine learning in general) via both practical experiments and theories with mild assumptions. Empirically, by proposing new algorithms and systems, we show that (active) model learning with strong priors on model structures makes it feasible to adopt advanced planners to solve complicated long-horizon robotic manipulation problems that were not possible before. On the other hand, we verify our theories through mathematical analyses of data efficiency for our active data acquisition strategies based on Bayesian optimization and systems combining learning and planning. The new approaches integrate structural prior knowledge with statistical machine learning methods to achieve state-ofthe- art performance on complex long-horizon robot manipulation tasks.