Learning to generalise grasps to novel objects

Add to your list(s) Download to your calendar using vCal

• Marek Kopicki, University of Birmingham
• Thursday 04 July 2013, 16:00-17:00
• G29, Mech Eng .

If you have a question about this talk, please contact .

Host: Aaron Sloman

Please note that coffee will be served at 3:15 at room 123 (School of Computer Science), followed by the one-hour seminar at 4pm.

Generalising grasps to novel objects is an open problem in robotics. In this talk I will present a method that can learn grasps for high degree of freedom robots that generalise to novel objects, given as little as one demonstrated grasp. The method is potentially more general and can be used not only in grasping, but also in any kind of robotic applications that involve robot body–environment contacts. The example could be dexterous manipulation, manipulation of deformable objects, walking, etc.

During grasp learning two types of probability density are learned that model the demonstrated grasp. The first density type (the contact model) models the relationship of an individual robot link to a local surface feature at its contact point. The second density type (the robot configuration model) models the whole robot configuration which is preferable for a particular grasp type. When presented with a new object, many candidate grasps are generated, and a grasp is selected that maximises the product of these densities.

The experimental results show successful grasp transfers to novel objects. The experiments include cases where the robot does not have a complete 3D model of the object being grasped, and grasps executed with two different multi-finger hands.

This talk is part of the Computer Science Departmental Series series.

This talk is included in these lists:

• Computer Science Departmental Series
• Computer Science Distinguished Seminars
• G29, Mech Eng
• computer sience

Note that ex-directory lists are not shown.