Reliable Manipulation

The seminar 'Reliable Manipulation' by Robert Howe (Harvard University) will take place on on Monday, April 28, at 10 a.m. in Room 1 of the BioRobotics Institute (Viale Rinaldo Piaggio 24, Pontedera). The event is the result of a collaboration between the BioRobotics Institute of the Sant'Anna School and Harvard University. The seminar can also be followed on Teams (link on the poster).

Biosketch 

Robert D. Howe is the Abbott and James Lawrence Professor of Engineering at the Harvard Paulson School of Engineering and Applied Sciences, and Founding Co-Director of the Harvard MS/MBA Degree Program. Dr. Howe started the Harvard BioRobotics Laboratory in 1990, which investigates the roles of sensing and mechanical design and motor control, in both humans and robots. His research interests focus on manipulation, the sense of touch, and human-machine interfaces. Biomedical applications of this work include of robotic and image-guided surgery. Dr. Howe earned a bachelor’s degree in physics from Reed College, then worked as a design engineer in the electronics industry in Silicon Valley. He received a doctoral degree in mechanical engineering from Stanford University in 1990, and then joined the faculty at Harvard. Dr. Howe is a Fellow of the IEEE and the AIMBE, and has received Best Paper Awards at mechanical engineering, robotics, and surgery conferences. 
Lab web site: http://biorobotics.harvard.edu/

Abstract 

Robots in unstructured environments like homes and workplaces must safely handle a vast range of objects. While recent robotics research has demonstrated increasing grasping success rates, it is not clear that current methods can approach the requirements of real-world applications. We begin by quantifying tolerable failure rates in unstructured settings, with the conclusion that dropped object rates must be less than 1 in 1,000 to 1 in 10,000 for many applications. To achieve this reliability, we are developing error detection mechanisms that can operate in parallel with robot grasping control systems. These systems can alert the controller when an object may be dropped, which allows the system to prevent or correct the error. One example is a collision detection algorithm that uses tactile signals from the robot fingers to estimate the location and force of a collision between a grasped object and a surface in the environment. This information can enable the controller to replan the task to avoid the collision, or it can be used to control contact during manipulation tasks. Another approach uses signals from the sensors in the robot hand to estimate whether a grasp will be stable under the anticipated forces in executing a task; if the fingers are projected to slip, then the controller can regrasp the object to guarantee stability. This system uses a hybrid of physical models (grasp analysis) and machine learning methods to maximize effectiveness. We conclude with a discussion of other potential error detection and correction mechanisms and the sensors needed to implement them.