BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Motor Skills Learning for Robotics - Jan Peters\, Max Planck Insti
 tute for Biological Cybernetics in Tübingen
DTSTART:20091009T120000Z
DTEND:20091009T130000Z
UID:TALK20882@talks.cam.ac.uk
CONTACT:Carl Edward Rasmussen
DESCRIPTION:Autonomous robots that can assist humans in situations of dail
 y life\nhave been a long standing vision of robotics\, artificial intellig
 ence\,\nand cognitive sciences. A first step towards this goal is to creat
 e\nrobots that can learn tasks triggered by environmental context or\nhigh
 er level instruction. However\, learning techniques have yet to\nlive up t
 o  this promise as only few methods manage to scale to\nhigh-dimensional m
 anipulator or humanoid robots. In this talk\, we\ninvestigate a general fr
 amework suitable for learning motor skills\nin robotics which is based on 
 the principles behind many analytical\nrobotics approaches. It involves ge
 nerating a representation of motor\nskills by parameterized motor primitiv
 e policies acting as building\nblocks of movement generation\, and a learn
 ed task execution module\nthat transforms these movements into motor comma
 nds.\nWe discuss learning on three different levels of abstraction\, i.e.\
 , learning\nfor accurate control is needed to execute\, learning of motor 
 primitives\nis needed to acquire simple movements\, and learning of the ta
 sk-dependent\n"hyperparameters" of these motor primitives allows learning 
 complex\ntasks. We discuss task-appropriate learning approaches for imitat
 ion\nlearning\, model learning and reinforcement learning for robots with\
 nmany degrees of freedom. Empirical evaluations on a several robot\nsystem
 s illustrate the effectiveness and applicability to learning control\non a
 n anthropomorphic robot arm.
LOCATION:Engineering Department\, CBL Room 438
END:VEVENT
END:VCALENDAR