BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//TUC//Events//EN
CALSCALE:GREGORIAN
BEGIN:VTIMEZONE
TZID:Europe/Athens
TZNAME:EEST
DTSTART:19700329T030000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=3
BEGIN:STANDARD
TZOFFSETFROM:+0200
TZOFFSETTO:+0300
TZNAME:EET
DTSTART:19701025T040000
RRULE:FREQ=YEARLY;BYDAY=-1SU;BYMONTH=10
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CREATED:20221110T115256Z
LAST-MODIFIED:20221110T115256Z
DTSTAMP:20260515T010825Z
UID:1778796505@tuc.gr
SUMMARY:Ομιλία κου Κωνσταντίνου Χατζηλυγερού
 δη "Micro-Data Reinforcement Learnin
 g for Adaptive Robots"
LOCATION:
DESCRIPTION:https://www.ece.tuc.gr/el/katalogos-
 ekdiloseon?tx_tucevents2_tuceventsdi
 splay%5Baction%5D=show&tx_tucevents2
 _tuceventsdisplay%5Bcontroller%5D=Ev
 ent&tx_tucevents2_tuceventsdisplay%5
 Bevent%5D=5885&cHash=95663bf014dda86
 d81879686306b1db2\nAbstract\n Robots
  have to face the real world, in whi
 ch trying something might take secon
 ds, hours, or even days. Unfortunate
 ly, the current state-of-the-art lea
 rning algorithms (e.g., deep learnin
 g) rely on the availability of very 
 large data sets. In this talk, we ex
 plore approaches that tackle the cha
 llenge of learning by trial and erro
 r in a few minutes on physical robot
 s (we call this challenge "micro-dat
 a reinforcement learning"). We will 
 discuss methods that utilize the sim
 ulator to generate behavior repertoi
 res in order to develop algorithms t
 hat allow complex robots to quickly 
 recover from unknown circumstances (
 e.g., damages or different terrain) 
 while completing their tasks and tak
 ing the environment into account. In
  particular, we will see how a physi
 cal damaged hexapod robot can recove
 r most of its locomotion abilities i
 n an environment with obstacles, and
  without any human intervention, usi
 ng this type of algorithms. On a sim
 ilar note, we will discuss how we ca
 n use different representations of p
 olicy functions to enable faster and
  more reliable behavior learning. In
  more detail, we will see how encodi
 ng the policy function as a trajecto
 ry can reduce the speed of learning 
 in robotic manipulation tasks, while
  also giving us theoretical guarante
 es about the behavior. Next, we will
  discuss how model-based reinforceme
 nt learning (RL) algorithms can be a
 dapted so that we can use them on re
 al physical robots. In particular, w
 e will discuss (1) methods that leve
 rage multi-core CPUs to enable fast 
 computational times, and (2) how we 
 can "scale" model-based RL methods t
 o high-dimensional robots. More conc
 retely, we will showcase algorithms 
 that are able to find high-performin
 g walking policies for a physical da
 maged hexapod robot (48D state and 1
 8D action space) in less than 1 minu
 te of interaction time. Next, we wil
 l present methods that aim to incorp
 orate learning methods inside tradit
 ional control architectures. In part
 icular, we will present work towards
  incorporating data-driven methods i
 nto QP-based controllers and showcas
 e how we can use these hybrid approa
 ches to achieve robust tracking and 
 performance on real-world robots and
  applications. Finally, we will disc
 uss current work towards autonomous 
 skill discovery and learning in robo
 tics applications.\n About the speak
 er\n Dr. Konstantinos Chatzilygeroud
 is received the Integrated Master de
 gree (Engineering Diploma) in comput
 er science and engineering from the 
 University of Patras, Patras, Greece
 , in 2014, and the Ph.D. degree in r
 obotics and machine learning from In
 ria Nancy-Grand Est, France and the 
 University of Lorraine, Nancy, Franc
 e in 2018. From 2018 to 2020 he was 
 a Postdoctoral Fellow with the LASA 
 Team with the Swiss Federal Institut
 e of Technology Lausanne (EPFL), Lau
 sanne, Switzerland. He is a recipien
 t of an H.F.R.I. Grant for Post-doct
 oral Fellows (2022-2024): he is the 
 Principal Investigator of the projec
 t "Novel Optimization Methods for Au
 tonomous Skill Learning in Robotics"
  that is being implemented within th
 e Department of Mathematics, Univers
 ity of Patras, Greece. He has also t
 aught and is still teaching several 
 undergraduate and post-graduate cour
 ses on Artificial Intelligence, Comp
 uter Science and Robotics at Univers
 ity of Patras, Greece. He has also c
 o-supervised several undergraduate a
 nd master theses. He is currently se
 rving as an Associate Co-Chair of th
 e IEEE Technical Committee on Model-
 based Optimization for Robotics, whi
 le he has served as an Associate Edi
 tor for several years at the Interna
 tional Conference on Intelligent Rob
 otics (IROS) and actively participat
 ed in the organization committee (as
  a Chair responsible for the virtual
  part of the conference) of the Inte
 rnational Conference on Robot Learni
 ng (CoRL) 2021. His work has been pu
 blished in top-tier journals and con
 ferences in the fields artificial in
 telligence, machine learning and rob
 otics, and he has received a Best Pa
 per Award at GECCO 2022. He has also
  actively collaborated with industri
 al partners: he was the Leader of th
 e R&D Computer Vision Team at Me
 targus, a pre-seed funded startup (b
 ased in Patras, Greece), and he was 
 the Lead Robotics Engineer at Ragdol
 l Dynamics (company based in London,
  UK). His research interests include
  the area of artificial intelligence
  and focus on reinforcement learning
 , fast robot adaptation, evolutionar
 y computation and autonomous skill d
 iscovery.\n
STATUS:CONFIRMED
ORGANIZER;RSVP=FALSE;CN=TUC;CUTYPE=TUC:mailto:webmaster@tuc.gr
DTSTART:20221110T160000
DTEND:20221110T170000
TRANSP:OPAQUE
CLASS:DEFAULT
END:VEVENT
END:VCALENDAR