State estimation and planning under uncertainty for robot manipulation

In the current approach to automated manufacturing, robots behind cages are programmed via tedious teach-in to perform repetitive tasks. This approach is becoming economically unviable as industrial automation transitions away from mass production in large batch sizes toward mass customization and ever-shorter product cycles. To ready robots for this change, two central bottlenecks must be addressed. First, tasks in small-batch automation are very diverse and change frequently; a task might even be executed only once. Therefore, robots must be able to solve the task at hand autonomously. Second, many tasks involve high degrees of uncertainty. In view of frequently changing tasks, it is no longer possible to eliminate uncertainty manually, as is currently done using fixtures and part feeders. Instead, a robot must be designed to systematically take into account uncertainty, weigh risks, and, if necessary, incorporate sensory feedback. This thesis addresses these two bottlenecks. We propose a holistic framework for general manipulation under uncertainty, in which a robot can systematically incorporate uncertainty within its decision-making process. As detailed below, we develop a universal model with algorithms for estimation, planning, and real-time control. Modeling: Our model for contact-rich manipulation under uncertainty is the first probabilistic model for manipulation that can efficiently represent multi-modal distributions of the full state – that is, for both positions and velocities. The model is based on an integral combination of contact dynamics and compliant, torque-based robot control. Estimation: Based on this model, we introduce a state estimator for contact-rich manipulation tasks. The estimator is capable of tracking the multi-modal distribution of the state of multiple robots and (even articulated) objects in real time. Such precise real-time estimates are an indispensable basis for controlled and reactive robot manipulation. Planning: To enable robust autonomous assembly from coarsely ...