UC Berkeley

Although robots play increasingly important roles in automated production due to their high efficiency, high accuracy, and high repeatability, new challenges for robots arise from different aspects as market demands shift and technology improves. The increasing product complexity and shorter product life cycle bring more difficulties into factories. Hence, higher intelligence for robots is necessary to perform various complicated tasks and to safely assist human workers.

In order to enhance robot intelligence, one could consider referencing the pattern of human development. When dealing with a completely new task, humans would learn or ask for assistance from experienced people or experts. Once learned, humans would apply such skill to various similar tasks. Furthermore, rather than merely completing the task, humans would make plans to accomplish the mission with better quality and efficiency. The following three phases could be referred to if we apply the same pattern to robot intelligence: 1) Programming, 2) Learning, and 3) Planning. Programming is to retrieve the information/knowledge from human. Learning is to generalize the learned skill to similar tasks. Planning is to plan an optimal policy to achieve the goal given constraints.

Following the aforementioned phases, this dissertation is divided into three parts to study the three phases. The programming part investigates several alternative programming approaches and introduces an online collision avoidance algorithm for human guidance programming. Following the framework of learning from demonstration, the learning part proposes remote lead through teaching (RLTT) for assembling and grinding skill learning and applies the non-rigid registration algorithm - coherent point drift (CPD) to transfer the learned grasp examples to similar objects. The planning part presents a fast robot motion planner by using the convex feasible set (CFS) to solve the nonconvex optimization problem in collision avoidance path planning and operational time reduction.