UC Berkeley

Industrial robot manipulators are widely deployed in various manufacturing tasks. Compared with human workers, industrial robot manipulators have advantages in terms of precision, efficiency, and repeatability. But it often requires tremendous engineering efforts to set up and program the manipulator for a specific task. The deficiency of intelligence restricts robots from broader applications. Therefore, it becomes more and more important to enable robots to acquire skills that can accomplish complex tasks and generalize across different scenarios. This dissertation aims to develop skill learning and planning methods for industrial robotic manipulation. We study 1) how to learn manipulation skills when there are uncertainties in the object state estimation, 2) how to generalize the manipulation skills across different scenarios, 3) how to achieve high-level task planning for long-horizon manipulation tasks.

Robotic manipulation of both rigid and deformable objects is studied in this dissertation. To manipulate rigid objects, a contact pose identification method is proposed to compensate for the pose uncertainties in the peg-in-hole assembly. In addition to rigid objects, the manipulation of deformable objects is also studied. A tracking and manipulation framework is proposed to robustly estimate the state of the cable and manipulate the cable to desired shapes. For more complex cable manipulation tasks, which often require long-horizon planning, a spatial representation is proposed to model the spatial relationship between the cable and environment fixtures. Multiple manipulation primitives are efficiently learned to configure the cable to desired states. For the task that combines both assembly and deformable object manipulation, a trajectory optimization with complementarity constraints is formulated to model the hybrid dynamics in belt drive units assembly. The problem is solved as a mathematical program with complementarity constraints to obtain feasible and efficient assembly trajectories.