BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.cam.ac.uk//v3//EN
BEGIN:VTIMEZONE
TZID:Europe/London
BEGIN:DAYLIGHT
TZOFFSETFROM:+0000
TZOFFSETTO:+0100
TZNAME:BST
DTSTART:19700329T010000
RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU
END:DAYLIGHT
BEGIN:STANDARD
TZOFFSETFROM:+0100
TZOFFSETTO:+0000
TZNAME:GMT
DTSTART:19701025T020000
RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU
END:STANDARD
END:VTIMEZONE
BEGIN:VEVENT
CATEGORIES:Machine Learning @ CUED
SUMMARY:Graph neural network approach for decentralized mu
 lti-robot coordination - University of Cambridge
DTSTART;TZID=Europe/London:20220421T110000
DTEND;TZID=Europe/London:20220421T120000
UID:TALK173051AThttp://talks.cam.ac.uk
URL:http://talks.cam.ac.uk/talk/index/173051
DESCRIPTION:Effective communication is key to successful\, dec
 entralized\,\nmulti-robot path planning. Yet\, it 
 is far from obvious what\ninformation is crucial t
 o the task at hand\, and how and when it must\nbe 
 shared among robots. To side-step these issues and
  move beyond\nhand-crafted heuristics\, we propose
  a combined model that\nautomatically synthesizes 
 local communication and decision-making\npolicies 
 for robots navigating in constrained workspaces. O
 ur\narchitecture is composed of a convolutional ne
 ural network (CNN) that\nextracts adequate feature
 s from local observations\, and a graph neural\nne
 twork (GNN) that communicates these features among
  robots. We train\nthe model to imitate an expert 
 algorithm\, and use the resulting model\nonline in
  decentralized planning involving only local commu
 nication\nand local observations. We evaluate our 
 method in simulations by\nnavigating teams of robo
 ts to their destinations in 2D cluttered\nworkspac
 es. We measure the success rates and sum of costs 
 over the\nplanned paths. The results show a perfor
 mance close to that of our\nexpert algorithm\, dem
 onstrating the validity of our approach. In\nparti
 cular\, we show our model's capability to generali
 ze to previously\nunseen cases (involving larger e
 nvironments and larger robot teams).\nIn Today's t
 alk\, the speaker will present his work from proof
  of\nconcept in simulation into the real-world rob
 otics systems toward the\nfully decentralized syst
 em.\n\n\n\nSpeaker's Bio:\n\nQingbiao Li is a fina
 l year PhD student at Prorok Lab in the Digital\nT
 echnology Group (DTG) at the University of Cambrid
 ge under the\nsupervision of Dr Amanda Prorok. Dur
 ing his PhD\, he focuses on\ncommunication-aware m
 otion planning for multi-robot coordination. He\ni
 s investigating Graph Neural Networks (GNN) to bui
 ld communication\nchannels for multi-agent and mul
 ti-robot systems so that they can\nlearn how to co
 mmunicate between each other explicitly. His resea
 rch\ncan be applied to mobility-on-demand systems\
 , automated warehouses\,\nand smart cities.\n\nPri
 or to joining Cambridge\, he was working in the Ha
 mlyn Centre at\nImperial College London\, founded 
 by Prof Guang-Zhong Yang and the Lord\nAra Darzi\,
  in the field of medical robotics and healthcare t
 o earn my\nMRes in Medical Robotics and Image-Guid
 ed Intervention. He was\nsupervised by Prof Daniel
  Elson for an eight-month research project\nabout 
 oxygen saturation (StO2) estimation\, and graduate
 d with a\ndistinction. He also held an MEng degree
  in Mechanical Engineering\nfrom the University of
  Edinburgh.\n\nHomepage:  http://qingbiaoli.github
 .io/
LOCATION:Hybrid meeting\, CBL seminar room\, and Zoom https
 ://eng-cam.zoom.us/j/87988930234?pwd=eHRGTHFtUE9UL
 0pJOWNtdDF2VDN2UT09
CONTACT:Dr H Ge
END:VEVENT
END:VCALENDAR