BEGIN:VCALENDAR
VERSION:1.0
BEGIN:VEVENT
CATEGORIES:MEETING
SUMMARY;ENCODING=QUOTED-PRINTABLE:CSLI Coglunch=3A Pieter Abbeel on =22Apprenticeship Learning with Applications in Autonomous Helicopter Flight and Robot Dogs=22
DESCRIPTION;ENCODING=QUOTED-PRINTABLE:Many problems in control have unknown=2C stochastic=2C and highly non-linear dynamics=2C and offer significant challenges to classical control methods=2E Some of the key difficulties in these problems are that =28i=29 It is often hard to write down=2C in closed form=2C a formal specification of the control task =28for example=2C what is the objective function for =22driving well=22=3F=29=2C =28ii=29 It is difficult to learn good control---as opposed to merely descriptive---models of the dynamics =28cf=2E the =22exploration problem=22 in reinforcement learning=29=2C and =28iii=29 It is expensive to find closed-loop controllers for high dimensional=2C highly stochastic domains=2E  In this talk=2C I will present formal results showing how these problems can be efficiently addressed in the apprenticeship learning setting=2C in which expert demonstrations of the task are available=2E  I will also present an application of our ideas to autonomous helicopter flight and four-legged locomotion=2E  Our helicopter results significantly extend the state of the art in helicopter control=2C and include the first successful completion of the following five aerobatic flight maneuvers=3A tic-toc=2C in-place forward flip and sideways roll=2C nose-in funnel=2C and tail-in funnel=2E Our resulting robot dog controller constitutes the state of the art in quadruped locomotion across rugged=2C previously unseen terrains=2E=0D=0A=0D=0APieter Abbeel is a PhD student in Prof=2E Andrew Ng=27s group at Stanford University=2E  His research interests include machine learning=2C robotics=2C and control=2E=0D=0A http://www-csli.stanford.edu/events/Coglunch/ 
LOCATION;ENCODING=QUOTED-PRINTABLE:Nora Suppes Hall=2C room 103 =28next to Cordura Hall=29
DTSTART:20071115T200000Z
DTEND:20071115T210000Z
END:VEVENT
END:VCALENDAR