Carnegie Mellon UniversityTHE ROBOTICS INSTITUTE

Thesis DefenseMikePhillips

Tuesday, April 14, 2015Newell Simon Hall 3305

2:00 p.m.

Maxim LikhachevChair

Siddhartha Srinivasa

Manuela Veloso

Thesis Committee

Experience Graphs: Leveraging Experience in Planning

AbstractMo#on  planning  is  a  central  problem  in  robo#cs  and  is  crucial  to  finding  paths  to  navigate  and  manipulate  safely  and  efficiently.  Ideally,   we   want  planners  which   find  paths  quickly   and  of  good  quality.   Addi#onally,   planners   should   generate   predictable  mo#ons,  which  are  safer  when  opera#ng  in  the  presence  of  humans.  While  the  world  is  dynamic,   there  are   large  parts  that  are  sta#c  much  of  the   #me.  For   instance,  most   of  a   kitchen   is  fixed  and  factory  floors  are   largely   sta#c  and  structured.   Further,  there   are   many   tasks  in   these   environments   that   are   highly   repe##ve.   Some   examples  are   moving   boxes  from  a   pallet   to  shelving  in  a  warehouse,  or  in  a  kitchen  when  moving  dirty  dishes  from  a  sink  to  dishwasher.  This  thesis  studies  how  to  exploit  the  repe##on  of  these  tasks  to  improve  planning  by  learning  from  past  experience  or  human  demonstra#ons.  

At  a  high  level,   the  proposed  planning  framework  takes  a  set  of  previous  plans  which  may  have  been  generated  by  the  planner  previously,  found  by  some  other  planner,  or  provided  from  a  human  demonstra#on.  These  prior  plans  are  put  together  to  form  an  Experience  Graph  or  EGraph.  When  solving  a  new  problem,  the  planner  is  biased  toward  parts  of  the  Experience  Graph  that  look  as  though  they  will   help  find  the   goal   faster.   Our   experiments  show   that  in  repe##ve   tasks,   using   E-­‐Graphs  can  lead  to  large   speedups  in   planning  #me.  This  is  done   in  a   way   that   can  provide   guarantees  on   completeness  and  the  quality   of  the  solu#ons   produced,   even   when   the   prior   experiences   have   arbitrary   quality   (for   instance,   when   based   on   a   human  demonstra#on).  

Experimentally,  we  have  applied  E-­‐Graphs  to  high  dimensional  pick-­‐and-­‐place   tasks  such  as  single-­‐arm  manipula#on  and  dual-­‐arm  mobile  manipula#on.  One  such  experiment  was  an  assembly  task  where  the  PR2  robot  constructed  real  birdhouses  out  of  wooden   pieces   and   nails.   We   also   applied   E-­‐Graphs   to  mobile   manipula#on   tasks   with   constraints,   such   as   approaching,  grasping,  and  opening  a   cabinet  or  drawer.  Most  of   these  experiments  have  been  duplicated  in  simula#on  and  on  a   real  PR2  robot.   Our   results  show   that  under   certain  condi#ons,  E-­‐Graphs  provide   significant  speedups  over  planning   from  scratch  and  that  the   generated  paths  are   consistent:  mo#ons  planned  for   similar   start  and  goal  states  produce  similar  paths.  Addi#onally,  our   experiments  show  E-­‐Graphs  can  incorporate   human   demonstra#ons  effec#vely,   providing  an   easy  way   of   bootstrapping  mo#on  planning  for  complex  tasks.

Sven KoenigUniversity of Southern California

Sachin ChittaSRI International