Students: Adam Barber

Postdocs: Ji-Chul Ryu

Professors: Kevin Lynch

Overview

Nonprehensile manipulation primitives such as rolling, sliding, pushing, and throwing are commonly used by humans but are often avoided by robots, who generally use grasping. Dynamic nonprehensile manipulation raises challenges in high-speed sensing and control, as the manipulated object is not in static equilibrium throughout the process which would be the case with standard grasping. An advantage, however, is that dynamics can be exploited to help the robot control object motions that would otherwise be impossible.

Our long-term goal is to develop a unified framework for planning and control of dynamic robotic manipulation.  A typical manipulation plan consists of a sequence of manipulation primitives chosen from a library of primitives, with each primitive equipped with its own feedback controller.  Problems of interest include planning the motion of the manipulator to achieve the desired motion of the object and feedback control to stabilize the desired trajectory. 

As a first step to understand the nature of dynamic nonprehensile manipulation, we study feedback stabilization of a canonical rolling example:  balancing a disk-shaped object on top of a disk-shaped manipulator in a vertical plane.  We constrain the motion of the manipulator to rotation about its center.  Using backstepping and assuming rolling contact at all times, we derive a control law that asymptotically stabilizes the object to the balanced position from any initial state.  The basin of attraction is necessarily reduced, but still large, when the contact is modeled according to Coulomb friction.  This work is successfully implemented using high-speed vision feedback.

We are currently focusing on motion planning and feedback stabilization of a broad class of rolling trajectories for smooth planar objects rolling on smooth planar hands (other than disks) moving with a full three degrees-of-freedom.

Video