The Robotics Institute

I work on manipulation. My goal is to enable robots to robustly and gracefully interact with the world to perform complex manipulation tasks in uncertain, unstructured, and cluttered environments. I want to make this interaction faster, safer, elegant, and involve simpler actuation. To this end, I founded and direct the Personal Robotics Lab, co-direct the Manipulation Lab, and lead the HERB effort of the QoLTbots systems area and the Mobile Manipulation effort of the Mobility and Manipulation Thrust at the Quality of Life Technologies NSF ERC.

I am currently focussing on two topics: Physics-based Manipulation and The Mathematics of Human-Robot Interaction. They are heavily intertwined, both born out of the goal of robots performing complex manipulation tasks with and around people.

Physics-based Manipulation:
I focus on using physics in the design of actions, algorithms, and hands for manipulation:

• Manipulation is more than pick-and-place. We are developing nonprehensile physics-based actions and algorithms to reconfigure clutter in the way of the primary task by pulling, pushing, sweeping, and sliding it out of the way. We have also recently shown that a class of tactile localization problems can be formulated to be submodular.
• I am particularly fond of functional gradient methods which have been used with much success in physics. We have developed CHOMP, a functional gradient optimizer for robot motion planning, and variants like GSCHOMP that exploit the structure of manipulation problems.
• I believe simple hands can do complex things. I am working on building robustness into the design and algorithms of simple hands, embracing physics and underactuation to stabilize objects.

The Mathematics of Human-Robot Interaction: I focus on formalizing Human-Robot Interaction principles using machine learning, motion planning, and function gradient algorithms:

• We have been working on enabling seamless and fluent human-robot handovers. We have developed a taxonomy of human and dog handovers, designed expressive grasps and motions, and used time-series analysis to learn the communication of intent. Our JHRI paper summarizes this work.
• We are formalizing assistive teleoperation, framing it as the joint tightly-coupled problem of prediction, solved with machine learning, and arbitration by policy blending, solved with control theory. We are also working on the online adaptation of teleoperation interfaces with kernel machines.
• Our latest work formalizes predictable and legible motion as inference problems, bringing together concepts in psychology, animation, and machine learning. We are presently working on generating legible motion via functional gradient optimization.

I have greatly enjoyed collaborating with HRI researchers, and our joint exploration has broadened my understanding of humans, robots, and mathematics.

I am also interested in Manipulation Planning:
extending randomized planners to constraint manifolds and Perception for Manipulation: my group has developed MOPED: an efficient object recognition and pose estimation system for manipulation.