Toward Robot Learning for Fine Manipulation

People may think robots are capable when asked their impressions. Indeed, at the moment one is reading, hundreds and thousands of industrial robots are moving heavy car chassis in automobile factories and palletizing big piles of packages in e-commerce warehouses. However, if robots are so capable and helpful, why do we see so few robotic applications outside of highly structured industrial environments, in domestic environments such as our homes and businesses?After decades of research and development, robots might have been more advantageous than humans in mechanical capacity and precision, but they are still far less intelligent. Robots programmed through traditional model-based engineering are still much in awe of the ubiquitous variations in unstructured domestic environments.We argue that robots' manipulation skills must be adaptive and actively compliant to reach human-level performance and enable their wide adoption in unstructured domestic environments. We illustrate our attempts to bring the gap closer between robots' and humans' manipulation skills in the work presented in this dissertation.First, we explain why we think learning-based approaches are promising in closing the skill gap and survey the state-of-the-art in learned robot manipulation. We categorize the challenges in unstructured domestic environments for robot manipulation and analyze various learning techniques used in existing literature through the lens of adaptability.We then attempt to push robot manipulation a step forward into domestic environments with cost and adaptability in mind. We introduce a novel, optimization-based resolved velocity controller to serve as the motion controller in manipulation tasks. The proposed velocity controller considers the robot's mechanical limits and allows the robot to behave optimally and safely in and around singular configurations.Next, we present the model-free reinforcement learning of a force controller on a robotic manipulation system built with a low-cost, information-poor tactile sensor. Despite the limited sensing capability, our system demonstrates the ability to perform non-prehensile, dexterous manipulation tasks without models of the manipulated objects and shows robustness to uncertain geometry.Finally, the resolved velocity controller and the learned force controller are combined in a parallel force and velocity control framework to learn fine manipulation primitive skills (such as box sliding and tilting with uncertainty in initial contact). Exploration and exploitation are essential for robots to learn in the training phase. We show that such behavior can also be learned and used after deployment to enhance the adaptability of fine manipulation skills. We highlight the challenges in learning fine manipulation skills and the effectiveness of our approach with two case studies. The learned fine manipulation primitive skills have good success rates in assigned tasks and demonstrate strong adaptability to variations in object and contact properties such as dimensions, weight, and contact friction coefficients.Our results suggest that affordable personal robots with fine manipulation skills and human-like adaptability can be developed in the future.