Title: Manipulators and Manipulation in High Dimensional Spaces
Advisor: Emo Todorov
Supervisory Committee: Emo Todorov (Chair), Mehran Mesbahi (GSR, A&A), Rajesh P.N. Rao, Dieter Fox, and Sergey Levine
Abstract: Recent technological advancements are aggressively pushing the state of the art in robotics, especially in the field of legged locomotion. Not only robust agile quadrupedal locomotion, but dynamic bipedal locomotion in an unstructured environment have been demonstrated. Despite the similarity of the problem between legged locomotion and hand manipulation -- both involve planning trajectories for kinematic trees interacting with the environment by means of contacts -- these advancements have not yet had a comparable impact on dexterous dynamic manipulation. Robotic manipulation still involves simple kinesthetic movements with simple low degree-of-freedom manipulators. Most real world solutions involve custom manipulators with task specific hand-tuned solutions. Given that the design goal of robotic devices is to interact with and manipulate the world so as to solve real tasks, having one custom manipulator with carefully tuned solution per task does not scale. If robotic devices are to perform tasks of daily necessities in home environments or dangerous tasks in hostile environments, robust and universal manipulators capable of handling a variety of tasks are required.
This talk will present a collection of results collectively pushing the state of the art towards achieving dynamic dexterous manipulation. In particular, I'll present the "ADROIT manipulation platform", which is a fast, strong and compliant tendon-driven system capable of driving a ShadowHand skeleton surpassing human capabilities. I will also elaborate on a real-time behavior synthesis framework capable of generating dynamic dexterous manipulation in high dimensional spaces. The framework generalizes well across different tasks as it only needs high-level task specifications as inputs in order to synthesize the details of the manipulation. I will conclude with results on transferring these behaviors using machine learning techniques to the ADROIT manipulation platform.