Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.

Why Robots may need to be self-­‐aware, before we can really trust them - Alan Winfield.

553 views

Published on

Alan Winfield's talk from AWASS 2013

Published in: Technology
  • Be the first to comment

  • Be the first to like this

Why Robots may need to be self-­‐aware, before we can really trust them - Alan Winfield.

  1. 1. Why  Robots  may  need  to  be  self-­‐aware,   before  we  can  really  trust  them   Alan  FT  Winfield   Bristol  Robo=cs  Laboratory   Awareness  Summer  School,  Lucca  26  June  2013  
  2. 2. Outline   •  The  safety  problem   •  The  central  proposi=on  of  this  talk   •  Introducing  Internal  Models  in  robo=cs   •  A  generic  Internal  Modelling  architecture,  for  safety   –  worked  example:  a  scenario  with  safety  hazards   •  Towards  an  ethical  robot   –  worked  example:  a  hazardous  scenario  with  a  human  and  a   robot   •  The  major  challenges   •  How  self-­‐aware  would  the  robot  be?   •  A  hint  of  neuroscien=fic  plausibility  
  3. 3. The  safety  problem   •  For  any  engineered  system  to  be  trusted,  it   must  be  safe   – We  already  have  many  examples  of  complex   engineered  systems  that  are  trusted;  passenger   airliners,  for  instance   – These  systems  are  trusted  because  they  are   designed,  built,  verified  and  operated  to  very   stringent  design  and  safety  standards     – The  same  will  need  to  apply  to  autonomous   systems    
  4. 4. The  safety  problem   •  The  problem  of  safe  autonomous  systems  in   unstructured  or  unpredictable  environments,  i.e.     –  robots  designed  to  share  human  workspaces  and   physically  interact  with  humans  must  be  safe,     –  yet  guaranteeing  safe  behaviour  is  extremely  difficult   because  the  robot’s  human-­‐centred  working   environment  is,  by  defini5on,  unpredictable     –  it  becomes  even  more  difficult  if  the  robot  is  also   capable  of  learning  or  adapta5on    
  5. 5. The  proposi=on   In  unknown  or  unpredictable  environments,  safety   cannot  be  achieved  without  self-­‐awareness  
  6. 6. What  is  an  internal  model?   •  It  is  an  internal  mechanism  for  represen=ng   both  the  system  itself  and  its  environment   – example:  a  robot  with  a  simula5on  of  itself  and  its   currently  perceived  environment,  inside  itself   •  The  mechanism  might  be  centralized,   distributed,  or  emergent   “..an  internal  model  allows  a  system  to  look  ahead  to  the  future   consequences  of  current  ac=ons,  without  actually  commiYng   itself  to  those  ac=ons”     John  Holland  (1992),  Complex  Adap=ve  Systems,  Daedalus.  
  7. 7. Using  internal  models   •  Internal  models  can  provide  a  minimal  level  of   func5onal  self-­‐awareness     – sufficient  to  allow  complex  systems  to  ask  what-­‐if   ques=ons  about  the  consequences  of  their  next   possible  ac=ons,  for  safety   •  Following  Dennea  an  internal  model  can   generate  and  test  what-­‐if  hypotheses:   –  what if I carry out action x..?! –  of several possible next actions xi, which should I choose?!
  8. 8. Dennea’s  Tower  of  Generate  and  Test   Darwinian  Creatures   Skinnerian  Creatures   Popperian   Creatures   Dennea,  D.  (1995).  Darwin’s  Dangerous  Idea,  London,  Penguin.   Natural   Selec=on   Individual   (Reinforcement)   Learning   Internal     Modelling  
  9. 9. Examples  1   •  A  robot  using  self-­‐ simula=on  to  plan  a   safe  route  with   incomplete  knowledge   Vaughan,  R.  T.  and  Zuluaga,  M.  (2006).  Use  your  illusion:  Sensorimotor  self-­‐  simula=on   allows  complex  agents  to  plan  with  incomplete  self-­‐knowledge,  in  Proceedings  of  the   Interna=onal  Conference  on  Simula=on  of  Adap=ve  Behaviour  (SAB),  pp.  298–309.  
  10. 10. Examples  2   •  A  robot  with  an  internal   model  that  can  learn   how  to  control  itself   Bongard,  J.,  Zykov,  V.,  Lipson,  H.  (2006)  Resilient  machines  through  con=nuous  self-­‐ modeling.  Science,  314:  1118-­‐1121.  
  11. 11. Examples  3   •  ECCE-­‐Robot   – A  robot  with  a   complex  body  uses   an  internal  model   as  a  ‘func=onal   imagina=on’   Marques,  H.  and  Holland,  O.  (2009).  Architectures  for  func=onal  imagina=on,  Neurocompu=ng  72,   4-­‐6,  pp.  743–759.   Diamond,  A.,  Knight,  R.,  Devereux,  D.  and  Holland,  O.  (2012).  Anthropomime=c  robots:  Concept,   construc=on  and  modelling,  Interna=onal  Journal  of  Advanced  Robo=c  Systems  9,  pp.  1–14.  
  12. 12. Examples  4   •  A  distributed  system  in   which  each  robot  has  an   internal  model  of  itself   and  the  whole  system   –  Robot  controllers  and  the   internal  simulator  are  co-­‐ evolved   O’Dowd  P,  Winfield  A  and  Studley  M  (2011),  The  Distributed  Co-­‐Evolu=on  of  an   Embodied  Simulator  and  Controller  for  Swarm  Robot  Behaviours,  in  Proc  IEEE/RSJ   Interna=onal  Conference  on  Intelligent  Robots  and  Systems  (IROS  2011),  San  Francisco,   September  2011.  
  13. 13. A  Generic  IM  Architecture  for  Safety   Internal  Model   Evaluates  the   consequences  of  each   possible  next  ac=on   Sense  data   Actuator     demands   The  loop  of   generate   and  test   The  IM  is  ini=alized   to  match  the  current   real  situa=on   Robot     Controller  The  IM   moderates   ac=on-­‐selec=on   in  the  controller   Copyright  ©  Alan  Winfield  2013  
  14. 14. A  Generic  IM  Architecture  for  Safety   Sense  data   Actuator     demands   The  loop  of   generate   and  test   Robot     Controller   Robot   Controller   Robot   Model   World   Model   Consequence   Evaluator   Object   Tracker  -­‐   Localiser   Copyright  ©  Alan  Winfield  2013  
  15. 15. N-­‐tuple  of  all   possible  ac=ons   (a1,  a2,  a3,  a4)   A  Generic  IM  Architecture  for  Safety   Sense  data   Actuator     demands   The  loop  of   generate   and  test   Robot     Controller   Robot   Controller   Robot   Model   World   Model   Consequence   Evaluator   Object   Tracker  -­‐   Localiser   Copyright  ©  Alan  Winfield  2013  
  16. 16. N-­‐tuple  of  all   possible  ac=ons   (a1,  a2,  a3,  a4)   A  Generic  IM  Architecture  for  Safety   Sense  data   Actuator     demands   The  loop  of   generate   and  test   Robot     Controller   Robot   Controller   Robot   Model   World   Model   Consequence   Evaluator   Object   Tracker  -­‐   Localiser   S-­‐tuple  of   safe  ac=ons   (a3,  a4)   Copyright  ©  Alan  Winfield  2013  
  17. 17. A  Generic  IM  Architecture  for  Safety   Sense  data   Actuator     demands   The  loop  of   generate   and  test   Robot     Controller   Robot   Controller   Robot   Model   World   Model   Consequence   Evaluator   Object   Tracker  -­‐   Localiser   S-­‐tuple  of   safe  ac=ons   (a3,  a4)   N-­‐tuple  of  all   possible  ac=ons   (a1,  a2,  a3,  a4)   Copyright  ©  Alan  Winfield  2013  
  18. 18.  A  scenario  with  safety  hazards   Consider  a  robot  that  has  four   possible  next  ac=ons:   1.  turn  leq   2.  move  ahead   3.  turn  right   4.  stand  s=ll  Hole   Robot Wall   Copyright  ©  Alan  Winfield  2013  
  19. 19.  A  scenario  with  safety  hazards   Consider  a  robot  that  has  four   possible  next  ac=ons:   1.  turn  leq   2.  move  ahead   3.  turn  right   4.  stand  s=ll   Hole   Wall   Robot   ac(on   Posi(on   change   Robot   outcome   Consequence   Ahead  leq   5  cm   Collision   Robot  collides  with  wall   Ahead   10  cm   Collision   Robot  falls  into  hole   Ahead  right   20  cm   No-­‐collision   Robot  safe   Stand  s=ll   0  cm   No-­‐collision   Robot  safe   Copyright  ©  Alan  Winfield  2013  
  20. 20. Towards  an  ethical  robot   Which  robot  ac=on  would  lead   to  the  least  harm  to  the  human?   Copyright  ©  Alan  Winfield  2013  
  21. 21. Towards  an  ethical  robot   Which  robot  ac=on  would  lead   to  the  least  harm  to  the  human?   Robot   ac(on   Robot   outcome   Human   outcome   Consequence   Ahead  leq   0   10   Robot  safe;  human  falls  into  hole   Ahead   10   10   Both  robot  and  human  fall  into  hole   Ahead  right   4   4   Robot  collides  with  human   Stand  s=ll   0   10   Robot  safe;  human  falls  into  hole   Outcome  scale  0:10,  equivalent  to  Completely  safe:  Very  dangerous   Copyright  ©  Alan  Winfield  2013  
  22. 22. Combining  safety  and  ethical  rules   IF for all robot actions, the human is equally safe! THEN (* default safe actions *)! !output s-tuple of safe actions! ELSE (* ethical actions *)! !output s-tuple of actions for least unsafe human outcomes! Consider  Asimov’s  1st  and  3rd  laws  of  robo=cs:   (1)  A  robot  may  not  injure  a  human  being  or,  through  inac=on,  allow  a  human   being  to  come  to  harm,     (3)  A  robot  must  protect  its  own  existence  as  long  as  such  protec=on  does  not   conflict  with  the  First  (or  Second)  Laws     Isaac  Asimov,  I,  ROBOT,  1950   Copyright  ©  Alan  Winfield  2013  
  23. 23. Extending  into  Adap=vity   Sense  data   Actuator     demands   The  loop  of   generate   and  test   Robot     Controller   Robot   Controller   Robot   Model   World   Model   Consequence   Evaluator   Object   Tracker  -­‐   Localiser   Learned/adap=ve   behaviours   Copyright  ©  Alan  Winfield  2013  
  24. 24. Extending  into  Adap=vity   Sense  data   Actuator     demands   The  loop  of   generate   and  test   Robot     Controller   Robot   Controller   Robot   Model   World   Model   Consequence   Evaluator   Object   Tracker  -­‐   Localiser   Learned/adap=ve   behaviours   Copyright  ©  Alan  Winfield  2013  
  25. 25. Challenges  and  open  ques=ons   •  Fidelity:  to  model  both  the  system  and  its   environment  with  sufficient  fidelity;     •  To  connect  the  IM  with  the  system’s  real   sensors  and  actuators  (or  equivalent);     •  Timing  and  data  flows:  to  synchronize  the   internal  model  with  both  changing  perceptual   data,  and  efferent  actuator  data;   •  Valida5on,  i.e.  of  the  consequence  rules.  
  26. 26. Major  challenges:  performance   •  Example  –  imagine   placing  this  Webots   simula=on  inside   each  NAO  robot:   Note  the  simulated  robot’s   eye  view  of  it’s  world  
  27. 27. A  science  of  simula=on:  the  CoSMoS   approach   The  Complex  Systems  Modelling  and  Simula=on  (CoSMoS)  process,  from  Susan   Stepney,  et  al,  Engineering  Simula=ons  as  Scien=fic  Instruments  —  a  paaern  language,   Springer,  in  prepara=on.   The  CoSMoS  Process  Version  0.1:  A  Process  for  the  Modelling  and  Simula=on  of   Complex  Systems,  Paul  S.  Andrews,  et  al,  Dept  of  Computer  Science,  University  of  York,   Number  YCS-­‐2010-­‐453    
  28. 28. Major  challenges:  =ming   •  When  and  how  oqen  do  we  need  to  ini=ate   the  generate-­‐and-­‐test-­‐loop  (IM  cycle)?   – Maybe  when  the  object  tracker  senses  a  nearby   object  star=ng  to  move..?   •  How  far  ahead  should  the  IM  simulate   – Let  us  call  this  =me  ts.  if  ts  is  too  short  the  IM  will   not  encounter  the  hazard;  too  long  will  slow  down   the  robot.   – Ideally  ts  and  its  upper  limit  should  be  adap=ve.  
  29. 29. How  self-­‐aware  would  this  robot  be?   •  The  robot  would  not  pass  the  mirror  test   – Haikkonen  (2007),  Reflec=ons  of  consciousness   •  However,  I  argue  this  robot  would  be   minimally  but  sufficiently  self-­‐aware  to  merit   the  label   – But  this  would  have  to  be  demonstrated  by  the   robot  behaving  in  interes5ng  ways,  that  were  not   pre-­‐programmed,  in  response  to  novel  situa5ons   – Valida=ng  any  claims  to  self-­‐awareness  would  be   very  challenging  
  30. 30. Some  neuroscien=fic  plausibility?   •  Libet’s  famous  experimental  result  showed  that   ini=a=on  of  ac=on  occurs  before  the  conscious   decision  to  make  take  that  ac=on   –  Libet,  B  (1985),  Unconscious  cerebral  Ini=a=ve  and  the  role  of   conscious  will  in  voluntary  ac=on,  Behavioral  and  Brain  Science,   8,  529-­‐539.     •  Although  controversial  there  appears  to  be  a   growing  body  of  opinion  toward  consciousness  as   a  mechanism  for  vetoing  ac=ons   –  Libet  coined  the  term:  free  won’t  
  31. 31. In  conclusion   •  I  strongly  suspect  that  self-­‐awareness  via   internal  models  might  prove  to  be  the  only   way  to  guarantee  safety  in  robots,  and  by   extension  autonomous  systems,  in  unknown   and  unpredictable  environments   – and  just  maybe  provide  ethical  behaviours  too   Thank  you!   Reference  for  the  work  of  this  talk:  Winfield  AFT,  Robots  with   Internal  Models:  A  Route  to  Self-­‐Aware  and  Hence  Safer  Robots,   accepted  for  The  Computer  AJer  Me,  eds.  Jeremy  Pia  and  Julia   Schaumeier,  Imperial  College  Press,  2013.    

×