UC Berkeley

In many industries, such as manufacturing and logistics, where the setting can be highly structured or simplified, robotic manipulators have been shown to improve safety, efficiency, and productivity through the automation of dangerous and repetitive tasks. As we continue to explore the use of robots in more dynamic scenarios with advanced tools or uncertain environments, reliance only on precise position control is no longer viable. Complex interactions require the robot to either plan in advance or adapt reactively to achieve success. Model-based approaches allow robots to be controlled in a constrained manner while predicting how they affect the world around them without requiring large datasets which may be difficult to obtain due to time or safety reasons.

This dissertation presents methods of modeling, planning and control for robotic manipulators to perform complex tasks while using limited data to improve performance. We examine three different applications which have their own unique set of challenges including hybrid dynamics, noisy measurements, human-robot interactions with partial knowledge on obstacles, and utilizing tools that rapidly degrade with usage. Our approaches are tested in simulation and validated in hardware experiments.