So  I'm  very  excited  to  introduce my  friend  slash  colleague  Professor  Tricia  and  Kim, who  is  an  associate  professor  at  UIUC. He  received  his  PhD  in  national  university. And  then  he  worked  in  working  with Samsung  Advanced  Institute  of  Technology  to  develop some  awesome  hidden  secret  humanoid  technology. And  then  he  just  quit  a  job  just  to participating  in  the  darpa  Robotics  Challenge  heat. And  at  that  time  he  worked with  Carnegie  Mellon  University. Then  he  sat  down  with  digital  research  to create  a  very  awesome,  interesting,  joyful  robots. I  met  him  when  actually  it  is  in  research. And  after  that,  he  moved  to  academia. He's  now,  as  I  mentioned, the  associate  professor  at  UIUC. Associate  Professor  at  UIUC. Today,  I  believe  he  will  show us  super  interesting  robots, a  lot  of  them  that  just  one. So  I  will  hand  this  over  to  Professor  Chan  Kim  here. Hello  everyone.  I'm  shall  I  actually have  slides  for  all  this  introduction. So  I'll  just  start. My  talk  title  is  for  this  kind  of  general  tog,  always, I'm  using  the  same  title  towards  human  friendly  robots, but  the  contents  are  always  different. I  will  talk  about  most  resent  reserved  in  this  talk. Yeah,  this  is  me  and  I just  want  to  have  this  one  to  show  you. There's  a  YouTube  channel  for  my  lab. I'm,  I'm  nodding  cool  videos  in  this  channel. So  my  research  goal, the  life  goal  actually  is  make  better  human-like  robots. And  because  it's  not just  a  humanoid  robot,  humanoid  shape, human  shape  can  go  where  people  go, and  also  can  do  what  people  do. This  is  not  my  thing. This  is  actually  Jedi  Pratt. The  HMC  is  researchers. Now  he's  with  other  company. He  said  this  one, I  like  this  a  lot. But  what  I  think  is  through  this  robotics, through  humanoid  robot  that  we  can  understand the  human  and  the  robotics  itself. We  can  understand  the  nature  better. And  actual  goal  is  we  want  to help  people  in  daily  life  using  this  robotic  technology. So  from  nine  is  I started  working  on this  bipedal  locomotion  for  a  long  time. And  at  that  time  we  didn't  have any  commercialized  like  the  robots. So  I  started,  these  are  all  your  time. Very  slow,  but  still, this  one  works  pretty  well. So  from  very  simple  designed  to  this, this  one  is  actually  very  similar  to asthma  or  some  other  robots  using  harmonic  drive and  pulley  system  to have  to  have  six  degrees  of  freedom  per  lakh. So  I've  been  working  on  this for  a  long  time  in  grad  school. And  in  Samsung  Electronics. I'm  pretty  sure  no  one  actually  have  seen  this  before. This  is  robot  from  2008, 2012  ish,  Samsung  Electronics  develop  this  video. This  robot,  including  many other  manipulation  and  perceptual  net,  everything. And  the  interesting  part  of  this  robot, this  one  is  not  using  the  bent  knee. All  the  robots  are  using  actually  bent  knee. This  one  is  using  straightening,  walking. So  it's  pretty  different  approach  at  that  time. There  are  some  papers  about  this  one, but  this  one  is  kind  of  gone  project  now. In  2012  in  Samsung  Electronics, I  really  wanted  to  participate  in the  darpa  Robotics  Challenge  with  this  robot, but  that  was  not  happen. So  I  joined  Carnegie  Mellon  University to  work  on  darpa  Robotics  Challenge  and  in, during  diaper  texture  Linji, this  online  hierarchical  approach  to  make the  bipedal  locomotion  is  kind of  completed  by  many  research  groups, not  just  by  the  same  new  group. So  after  this,  we  I  was in  Christakis  and  Tim  WPI,  CMU  team. We  actually  codon  one, codon  when  the  game, but  the,  our  robot  is the  only  robot  didn't  use  intervention. So  we  think  the  controller  was  pretty  good. And  after,  during  this  TRC, I  had  a  chance  to  join  Disney. Disney,  I  spent  six  years to  develop  some  robotics  technology. To  make  some  features  for  the  animation  characters. So  there  are  many, many  features  in specific  animation  character  who  can  help, trigger,  can  hop,  and  there are  lots  sober  and  electronics  in  Disney. We  can  use  retargeting  methods  to  make natural  motion  for the  animation  character  and  Medtronic's. So  I  spent  some  time  to  develop 3D  printed  software  skins  for this  kind  of  character or  develop  some  other  new  type  of  actuator. This  one  is  named  as  lip  or  using  voice  coil and  the  parallel  elastic  component  that springs  to  make  this  a  hopping  motion. And  also  worked  on  the  retargeting, retargeting,  retargeting  problem  is really  interesting  because  the, if  the  source  data  source  or animation  or  motion  capture  data, we  want  to  use  this  data  to  different  configuration. We  need  to  solve  some  dimension  reduction  problem for  a  specific  robots. And  the  scale  problem  that  I  think graphics  people  already  explored  this  problem  a  lot, but  the  robot  has  its  own  constraints, like  speedier  or  collision  kind  of  problem. So  this  is  still  very  interesting  topic. And  then  2020,  right  before  COVID  Adrian, Yeah,  you  see,  to build  my  own  lab  and  to  work  on  some  interesting  robots. Lab  is  named  as  kidnap. It's  not  just  my  last  name  is kinetic  intelligent  machine,  which  is  robot. So  the  robot  in  lab. So  as  you  saw, what  I'm  doing  is  I'm  building  robots  from scratch  and  making  controller for  that  and  doing  some  interaction. I  said,  I  think  it's including  most  area  of  the  Robotics  field. But  I  can  say  what  I'm  doing  is  a  test  database, robot  design  and  control. So  if  there's  any  target  task  like  locomotion, bipedal  locomotion  or  interaction  hugging. We  collected  data  from  forward  that  task  and the  design  specific  mechanism  or  some  other  new  material using  some  other  new  material  to  do that  task  and  running  some  user  study or  experiment  to  collect  the  data and  improve  the  design  and  control  again  and  again. In  my  research  life, I've  been  working  on  many  robots. And  these  yellow  boxed  ones are  the  robots  are  built  from  scratch. And  the  red  ones  are  the  robust  legged  robots. So  I  can  say,  I've  been  working on  this  legged  robot  for  a  long  time. So  I  will  start  from  the  pipette  robot  first. So  pipette  robot  is  very  simple. If  a  robot  has  two  legs, we  can  call  that  one  bipedal  robot. And  the  history  of  this  bipedal  robot,  humanoid  robot, is  around  the  50  years  ago, started  from  this  whiteboard  from  Washington  University. And  all  the  robots  up  barriers  are  from Japan  and  in  us,  the  MacRae, but  the  Boston  Dynamics  founder  started  this  one  and  I  go hopper  and  there  are some  series  of  the  legged  robots  from  the  Leg  Lab. In  '90s. From  2000  when  Honda  announce  this esmolol  lasso  robots  or  have  been  developed  until  now. We  Central,  what's  our  Boston  Dynamics  atlas  or  Tesla, Optimus  and  Charmaz,  cyber  one. Those  are  bipedal  robots. And  these  are  some  examples  of  actually  the  robots  a  lot. And  all  these  robots  have  this  hip, knee,  and  ankle  joint. And  there  are  some  other  structure, some  other  mechanism,  but these  robots  are  using  revolute  joints. And  there  are  segment  linkages, pelvis,  thigh  and  calf  and  the  foot  links. And  these  links  are  all  rigid  links  and  robot. This  robot  is  standalone. Accommodate  with  robots,  so  it's  a  floating  body. This  is,  these  are very  obvious  fact  for  the  to  describe  this  bipedal  robot, but  these  are  really  important  assumption. We  need  to  have  two  more  Dell, dell  robot  and  run  the  simulation. Then  we  also  need  to  know  what  start  walking,  walking. These  are  from  the  dictionary. You  can  have  this  simple  definition  of  walking. Move  at  a  regular  or  fairly  slow paced  by  lifting  and  setting  on  each  foot  in  ton, never  having  both  feet  off  the  ground,  the  other  ones. So  if  there's  no  error, time  is  walking  is  way  too  simple. And  there  are  lots  of  synonyms  because  of  that. So  when  we  have  very  simple  definition, you  might  think  it's  easy  to  implement. It's  actually  harder  because some  people  can  think  this  one  is  working  but,  but, but  actual  walking  definition  of  walking in  everyone's  mind  is  different because  the  definition  is  simple. But  interestingly,  while  King  has been  studied  in  many  different  fields, and  this  one  is  from  medical  science. In  medical  science,  normal  walking was  studied  for  a  long  time  and  it's very  well  and  analyzed  with older  temporal  and  spatial  information. And  also  based  on  the  timing, the  function  on  the  T-cell, so  well  studied  it. So  researchers  are  working  on  this  field, is  using  this  kind  of  information  a  lot  to  design their  own  controller  and  how  to  make  this  walking, how  to  makeup  iPad  robot  walk  is the  main  research  topic  in  this  field. And  now,  as  I  said, after  DR.  securing  DRC, this  community,  the  kind of  merge  it  towards  one  direction. And  now  we  have  a  standard  process  for  bipedal  walking. Bipedal  king  is  basically  locomotion. So  there's  our  starting  point  and  the  goal  point. And  we  need  some  kind  of planner  first  to  have  this  cell  for  steps. So  first  step  planning  is  the  first  step or  to  design  the  walking  algorithm. So  you  cannot  just  have  random  files  tab, there  is  a  constraint  to  have  this. The  stem  lengths  or tiny  angle  based  on  your  hardware  or  your  model. You  need  to  have  that  kind  of information  to  actually  have  the  first  step. First. Then  based  on  the  first  step, when  you  have  first  tab, you  know,  when  the  foot  is  on  the  ground. So  based  on  that  information, you  can  have  target  center  of pressure  on  the  ground  at  the  right  timing. So  you  can  have this  targets  and  center  of  pressure  trajectory. You  need  to  generate  data  based  on  the  four-step  planner. And  based  on  that, the  first  step  planning and  COP  generation  is  still  underground. So  it's  not  in  3D  or  is  set  in  3D  actually, but  it's  not  using  the  3D  model  of  the  robot. So  researchers  in  this  field use  a  very  simple  linear  inverted  pendulum, or  the  spring  loaded  inverted  pendulum, or  this  table  model  to  use  the  swan  mass, one  mass  or  there  are  Martinez  model  but mostly  one  mass  to  find  the  trajectory  COM, trajectory  with  this  COP  information  on  the  ground. So  actually  solving  the  dynamics  to  satisfy this  COP  constraints  are, and  using  that,  you  can  have  this  purple  in  3D  space. So  after  that,  now  we have  footstep  and  CRM  movement  in  3D. We  can  use  all  this  information  to  get the  joint  level  commands  to  make  the  robot  walk. This  one  is  not  like  feedback  controller. This,  this  is  more  about how  to  generate  the  walking  motion. And  during  DRC,  this, uh,  optimization  based  semester. So. Not  using  all  the  models  analytic  solution, we  can  use  iteration  to  get closure  solution  for  a  good  solution  for  each  layer. And  we  could  solve  this  problem  pretty  well. But  the  meaning  of  this  processes, when  you  think  about this  dimension  for  the  footstep,  planning, for  step  is  maximum  is  one  for a  one-step  position  and  orientation. So  max,  like  six-degree  of  freedom  problem  here. And  after  that,  it's  almost  same  depends on  CRM  trajectory  is  based  on  very  simplified  model. So  the  dimension  in  this  space  is  pretty  low. And  using  the  model  and  all  the  information  together, we  can  actually  solve  this  high-dimensional  problem. High-dimensional  problem  is  meaning  that it  feels  so  this  humanoid  robot, there  are  like  30  motors. So  if  we  want  to  solve  the  dynamics  for  30  motor  system, you  will  need  the  velocity  and acceleration  and  probably  you  will need  this  some  kind  of fixed  the  coordinate  in  the  system. So  it  can  be  100  dimensional  problem. But  through  this  hierarchical  approach, you  can  make  it  in  very  low  dimensional  problem  and actually  get  this  solution  for  the  walking  motion. So  this  is  kind  of  standard  process  for  bipedal  walking. I  wanted  to  show  this  on  first to  talk  about  the  later  topic. And  I  will  just  say  introduced  one  robot, seven  and  I  worked  together  before  snap, but  it's  also  legged  robot. Snippet  looks  like  a  normal  hexapod, but  this  one  can  actually  lose  the  arm, the  limbs  like  that, and  can  change  its  locomotion  style based  on  the  current  configuration. So  this  is  a  very  interesting  design and  there  were  multiple  motivations  for  this. But  I  was  Disney  at  the  time. So  the  character's  name. What  is  the  feature  for  Olaf? So  all  love  can  survive  after  changes  body  shapes. So  Olaf  can  lose  body  part  and  can do  locomotion  or  whatever the  movement  based  on  the  current  configuration. So  that  was  one  good  motivation  in  Disney. But  the  really  interesting  part  to  us  at  this, sorry,  sorry  about  the  creepy. So  this  is  steady, long-neck,  long  leg  is  on  spider. And  this  one  is  actually  Poland losing  its  lag  when  there's  a  danger. And  after  losing,  after  changing this  configuration  is  actually  moving almost  same  speed  as  this  is  very  interesting because  it  was  born  with  n  lags, but  it  can  move  or  almost, in,  almost  same  speed  even  with  five  legs. So  this  is  really  interesting part  because  in  robotics  research, if  we  have  very  small  like  failure  in  the  system, we  cannot  use  than  robot  at  all. Fail-safe  system  is  really  important if  you  want  to  use  this  on your  word  and  how  to  handle  that  kind of  situation  is  very  important. I  will  show  you  the  snap  borders  developed  in  Disney. But  there  was,  I  wanted  to  add some  other  features  in  this  snapshot. So  developed  another  second  version  of  a  snap  or  in  UIUC. So  the  mechanism  is  very  simple. So  now  there  are  many  commercialize  the  motors. You  can  use  just  the  four  wires  or  three  wires to  power  the  motor  and communicate  with  the  motor  through  the  serial  connection. So  we're  using  that  kind  of sabo  motor  to  control  this  two  degree  of  freedom, lag  and  all  the  power, the  spring  loaded  power  kind  of spring-loaded  pogo  pin  was  used  for  the  connection. And  there  are  multiple  magnets for  the  hardware  connection  here. So  coupling  were  done  by  that  magnetic  force. And  the  older  batteries  and the  processors  and  the  camera  part  are  in  the  body  side. So  previous  snap  board could  do  this  configuration  recognition, which  is  really  important  in  this  system, using  just  a  current, which  was  not  the  best  method for  the  configuration  recognition. So  now  we  are  using  camera to  check  which  leg  is  connected  to  each  port. So  this  kind  of  motors  have  its  own  ID. So  when  you  ping  all  the  ID, you  can  tell  how  many  legs  are  connected. But  still  it's  ports  are  used in  current  configuration  is you  need  to  actually  check  that. So  if  you  have  a  baby  or  if  you  send  baby, babies  are  playing  with  their  hands  and the  feet  for  longtime  when  they're  really  young. That's  four  to  recognize their  current  config  configuration is  pretty  good  word  that  the  hands  and  feet, sorry,  there's  where  the  snapshot  is  developed  like  that. And  this  is  totally  open  source. Our  explain  this  later. But  for  study  on  legged  locomotion. And  this  is  where  we  worked  together  in  2008,  17,  18. And  when  I  saw  this  one, he  was  super  interested  in using  just  harder  for  the  reinforcement  learning. As  far  as  I  know,  this  is  almost  the  first  work  just  to using  hardware  for the  reinforcement  learning  using  one  camera. And  the  key  idea  was  actually  initializing system  to  make  the  robot  go back  to  the  initial  point  and  just  using  the  camera  to know  the  walking  distance  and  the  how, how  straight  or  king  was. So,  so  that  was  the  part  of  the  reward  function  for  this. And  this  is  a  very  interesting  work. And  I've  been  working on  distraction  with  some  other  collaborators. Because  after  this,  psi  got  lots  of  request from  machine-learning  field  to  share  that  design. So  now  I  made  this  second  version  as  open-source. And  if  you  check  the, this  is  the  GitHub  for  this  snippet,  a  second  version. And  you  can  also  find  this  cool  design. This  one  has  a  camera  inside  the, if  you  know  the  Toy  Story,  that  old  version, the  first  one,  you  can  tell this  one  is  from  the  Toystory. So  check  the  time. So  so  far. Well,  for  almost  15  years I've  been  working  on  this  legged  locomotion, developed  the  system  and implementing  the  controller,  a  better  controller. And  walking  is  really  interesting. Behavior. One  important  and  unique  human behavior  and  safety  critical  behaviors. So  that's  why  many  researchers  are  working  on  this. But  it's  a  well  explored  in  various  fields  and there  are  plenty  of  data actually  interesting  part  of  walking. But  what  we  wanna  do  is  not  just  the  walking. What  do  we  want  to  do  with  dropout? As  there  are  many  more  task  Nan, many  more  behaviors  we  can  do  with  robust  system. And  I  want  to, I  wanted  to  tackle  all  these  behaviors  or  task. But  it's  really  hard  to  find  some  good  general  solution. And  data  for  all  this  task  is  very  limited. It's  not  enough. And  I  started  thinking  like  robots  will, can  interact  with  the  environment and  objects  and  other  agents. So  we  can  just  start  collecting data  in  your  word  for  various  task. And  that  concludes  onwards, we  need  more  robots  to  solve this  practical  task  and  to  collect  real  world  data. So  I  kinda  moved  on  to  manipulation. Tesco  at  home. Robotics  manipulation  has  actually started  from  the  robotics. The,  the,  as  the  study  I  think. But  the  dexterous  manipulation  is  still  open  problem. And  industrial  robots  are  doing  all  the  task  very  well. In  the  constraint  environment  is  very  fast,  very  strong. But  using  that  large  robot  at  home  is  not  easy. And  how  to  use  this  manipulate,  robotic  manipulation. Tom  is  actually  very  difficult  problem. Lots  of  researchers  are  currently  working  on  that. And  there  are  many  systems  automate this  interesting  robot  and  part  robotics is  this  from  here, I  think.  Hello  robot. And  this  is  really  good  solution because  this  one  has  small  base  like  this. I  like  this,  a  low  libido  lot. Google's  everyday  robot. Samsung  and  other  robots  are. There  are  many  robots  now. Still  the  market  is  not  there. I  will  talk  about this  robot  vacuum  cleaner  a  little  bit  because a  Roomba  robot  vacuum  cleaner is  the  only  successful  Robot  application  for  home. Because  affordable  and  said, now  it's  a  much  more  expensive  I  think, but  the  cell  affordable  and can  do  the  one  important  task, cleaning  and  users  know  what  they  can  do. Now,  that  took  some  time,  I  think. But  most  important  part  is really  well  designed  for  human  environment. About  the  usage. Everyone  had  some  bad, like  memory  about  the  robot  vacuum. But  after  some  time, people  understood  how  it  works. So  we  can  actually  use  this  one  as  a  tool. Because  that  history  of  vacuum  cleaners  really interesting  because  I  thought that's  50  years  old  or  60  years  old, not  that  old,  but  it  started  from  Something  18,  69. This  mechanism  is  really interesting  because  if  you  appreciate  this  one, is  actually  moving  this  pump  to  clean  the  carpet. And  this  is  the  first  mechanism  as  the  vacuum  system. And  this  one  is  also  hand-held  vacuum  system  like  this. But  the  important  design  for  this  vacuum  machine, that  most  of  this  vacuum  machine  started  using this  wide  thin  section  part  like  this. And  over  past  hundred  years, the  human  environment  actually  hasn't  been  changed  it. So  when  we  purchase  some  kind  of  sofa  or  bad, we  are  now  considering  if  we  can  use vacuum  machine  underneath  this  furniture  so  that  spin, change  it  together  with  the  vacuum  cleaners. That's  a  really  interesting  part  to  me. And  like  all  the  companies, I  showed  companion  robots, I  showed  the  target  environment  for me  is  also  this  home  kitchen. Kitchen,  the  structure  1,980.19,  40,  60. Those  are  like  the  structure. The  design  wise  is  almost  similar. After  the  1960s, only  one  home  appliance was  added  next  to  some  error  here, this  dishwasher  and  this  home  kitchen  is  most structured  and  most  standardized  space  at  home, and  also  relatively  easy  to  modify. So  the  goal  is  simply  dishwashing is  simple  manipulation  where it  says  still  very  difficult  problem. And  many  researchers  wanted  to  solve. Idea  was  simple. I  want  it  to  have  some  error  here  to  do  this. And  I  don't  have  religion  to  be  clear, but  this  one  is  a  quinine, one  of  the  Buddha  and  I, there's  a  guanine  was  interested, interesting  to  me  because. Using  thousands  arms  to  help  people  in,  in, uh,  the  reason  why  it, she  has  the  thousand  times  is  to  help  people. And  I  kind  of  inspired  by  this and  made  this  kind  of  system  so  like  the  step. But  if  he  can  plug  the  robotic  arm  to  some  error. If  the  robot  can  do some  tests  in  the  past  to  location  for  that  task. I  think  we  can  use  this  by  modifying environments  slightly  as  just  a  ten  centimeter  get  there. So  the  current  problem  of this  manipulation  as  the  manipulator  arm is  way  too  expensive, even  for  the  research  use, are  easily  over  $20,000. And  the  older  companies  are working  on  this  mobile  manipulation  because  we need  like  multiple  arms  if  we need  to  use  that  one  as  a  fixed  feature. But  the  locations  at  home, the  kitchen  or  table, the  hours  and  the  frequency  of the  utilization  is  actually  pretty  short. So  so  the  design  goal was  making  this  pluggable  home  appliance. So  it's  not  about  developing. The  manipulator  is  more about  making  some  kind  of  standard, like  a  USB  port. We  can  power  the  robotic  arm  and  communicate  this with  this  system  to  do  some  task  in  proper  location. That  was  kind  of  concept. Further  design. And  current  version  has six  degrees  of  freedom  and  the  gripper  and  one  RGB-D. Intel. Your  sense  camera  is  I  touch  it. The  important  part  is  really  lightweight, under  six  kilogram  and  the  payload  is  still  about  3  kg. So  the  mechanism  or the  inside  mechanism  is actually  similar  to  the  snap  both. That's  why  I  wanted  to  show  the  Snapple  first. So  using  our  20  pins  for  the  power  and  the  communication, motor  communication  and  the  USB  communication. And  it  can  be  carried  simply  like  this. There's  some  small  design  features we  can  need  to  cover  the  important  part and  there's  a  locking  mechanism for  the  portable  probability. And  use  it  to  plug  like  this. Of  course,  you  need  to  change this  environment  slightly  because we  need  to  install  this  part. And  there's  a  power  and one  computer  for  the  motion  controller. But  still  just  a  four-inch  can  centimeter gap  somewhere  where  we  want  to  use  disarm. So  I  think  that  a  small  amount  of  modification  is  okay. For  this. We  also  check  the  workspace  to  check  the  payload. So  this  one  is  holding five-pound  dumbbell  to  test  the  payload. And  currently  the  software  stacks  are  all  based  on  rows. So  we  can  use  all  the  rows  packages  for  the  perception and  then  the  manipulation  planning every  layer  which  make  the, make  the  more  application  in  very  short  time. So  this  is, we  submitted  this  paper  and  is  currently  in  archives. So  if  you  are  interested  in  the  structure, you  can  check  that  paper. And  as  I  said, the  interface  with  other  devices  as  the  controller  input, you  can  simply  use  the  server  ER  or just  some  other  skeleton  estimation  functions  to, if  you  can  get. I'm  joined  today  by  command  from  somewhere  else. You  can  easily  just  to  use that  interface  to  control  the  arm. Well,  this  is  low-level  country, so  this  is  mostly  basic  rose  function. So  I  will  skip  this  part. I  will  just  show  you what's  the  current  applications  in  my  lab. So  this  is  papyrus  table. We  have  three  port  in  this  table. It  simple  coffee  dripping  them  all. So  how  to  make  this  kind  of motion  or  how  to have  this  solution  is  my  current  interests. I  believe  there  are  now  many  solutions from  language  to  robot  or  language  it  to  some  action. I  am  expecting  to  use  that  kind  of  method  a  lot. And  this  is  the  simple  form. We  have  many  arms, so  we  are  using  Ross, so  distributed  control  is  simple. So  if  we  want  to  do  the  synchronous,  nice, it's  the  not  actually  fully  synchronized, but  we  can  just  grab  that  topic  to  control  the  robot. And  making  mobile  robot  is also  very  simple  using  this  form. So  all  starts  says  double-headed  dark  and  we  are using  spot  and  there  is separate  bacteria  and  one  computer to  control  this  DRM  system. Making  this  spot  as  Durham  mobile  manipulators  simple. The  good  part  is  we  can actually  use  our  spouse  or  other  robots. Robot  your  base  robots  or  some  other  functions. Just  stay  in  our  system, not  having  many  other  sensors  in  the  upper  body  side. We  can  just  use  this  API  to  use  this, the  oldest  sensory  information. Actually,  this  one  is  doing everything  autonomously  with  some  script. But  what  we're  doing  is  just  so  from  some, from  some  point,  using  a  pretest, which  is  the  function  of  the  spot. We're  just  using  that  function for  our  mobile  manipulation. What  we're  doing  is  just two  upper  body  manipulation  side. So  using  some  other  devices  are  very  simple. And  this  is  slightly  weird  application, but  I  can  explain  this. So  I  wanted  to  make, frankly  i  0  need  to  make  the  IMS  setup  like  system. But  this  is  not  for  just  the  movie, is  if  you  think, if  we  will  use  the  wearable  robots  a lot  in  the  future  in  manufacturing  environment. Upper  body  for,  especially  for  upper  body. Wherever  robot,  someone  always  need to  help  the  user  to  where  the  system, how  to  make  that  on  as autonomous  system  will  be interesting  problem  isn't  a  good  excuse. So  I  wanted  to  make  this  system and  actually  made  this  up. Okay. So  this  one  was  last  year  at  Christmas  video. So  as  you  can  see, if  there's  a  target  wearable  device, such  as  a  kid. Putting  it  on  is  easy  because  the  subject  can  do  this. This  part  is  hard  because  because  we  need  to actually  control  the  robot  for  the  safer  interaction. So  currently  it's  just,  uh,  using  mock-up. But  interactive. This  syrup  and  synapses  term  will be  very  interesting  problem.  In  my  point  of  view. Well,  well. Welcome  to  Math.  And  so how  to  create  this  application  is  very  simple. So  sassy,  you  know,  UH, the  engineering  of  pronounced  in  the  UIUC, the  event,  big  event. So,  UH,  is  a  big  event  in UIUC  and  I  think  two  weeks  before, UH,  we  wanted  to  have  live  demo  with  some  robot. So  we  brought  this, but  also  wanted  to  show  some  cool  stuff  together. So  we  got  our  t  table  from  ikea  and  spent two  days  to  make  this  one  working  in  simulation. And  spend  three  days  to  attach  Chess  engine, which  was  not  developed  in  my  lab. So  we  actually  made  this  robot  arm  has to  play  and  do  this  task  in  front  of  this  audience. So  built  everything  in a  week  or  two  to  live  demo  for  two  days. So  if  you  have  3D  model, you  can  easily  just  modify  the  3D  CAD. And  the  planning  control  group  is  all  based  on  the, the  robots  and  we  already  have  all  these  things. So  the  attachment  point, though,  location  and  orientation  is  important. But  once  you  set  all  these  things, you  can  easily  do  any  test,  not  any  task, but  you  can  control  the  eye. What  I  want  to  talk  about  this,  Hi, I'm  trying  to  introduce  simple  concept. This  is  my  current  research  topic. There's  a  task  retargeting. What  I  wanna  do  is  given multimodal  data  for  specific  target  task, we  want  to  find  what  is  the  optimal  configuration, robot  configuration  for  that  test. And  we  want  to  solve  that  task  with  this. So  e.g.  coffee  dripping,  use. So  the  code  is  simply  High-level  by sequential  moving  something  somewhere  or  doing  some, some  motion  for  the  coffee  dripping. And  the  important  date  information and  stuck  configuration. We  need  to  design  that  part  for  that  task. For  this  task  actually. And  all  the  sequential  information  for that  task  is  actually  already  given  some  error. So  how  to  grab  all  this  information  to  make  this, this  sequence  is  important  and  how  to  solve  how  to  have this  pipeline  using  this  system  is  important. In  this  direction,  I  don  t  have  solution. This  is  my  current  research,  opaque. But  the  challenges  are, I  think  these  are  challenges  are  overcome. Configuration  that  workspace  is  really  important. Skills  for  the  task  skills  is more  related  to  reactive  control  sides. So  the  hardware  sensing  and its  interaction  between  object  tender  robot or  even  two  and  the  robot. How  to  make  that  skills. How  to  make  that  interaction  with  the  object is  important  than  the  experiment  is  important. The  first  one  can  be  solved  with the  weird  part  process  or  some  other  similar  system. I  I'm  not  sure. I'm  not  here  to  sell  my  system. What  I  wanna  do  is  I  want  to  share this  idea  to  use  this  kind  of  system, system  with  even  different  arms. So  the  arms  are  kind  of  bolted  down  on  top  of  your  lap. And  there  we  cannot  move  that  on  from there  to  some  other  place  in  a  day. So  what  I  wanna  do  is  just  if  we  have, if  we  know  what's  the  task? The  dishwashing,  Tesco's  moving  lots  of water  in  the  sink  for  up  from  sync  to  the  dishwasher. So  we  can  have  that  kind  of  data from  some  egocentric  like  vision  data  or  just so  you  can  collect  the  data  where  you  can heuristic  approach  to  figure out  where's  the  optima  installation  point for  that  test  to  solve  that  task? This  is  the  table  environment  there. So  this  space  is the  workspace  overlap  between  these  three  arms. Or  if  you  have  target  wearable  device,  you  can  actually. Optimize  the  number  of arms  and  location  and  orientation  and volumes  by  just  analyzing  the  process  for  the  syrup. And  yeah,  you  can  solve  many, many  task  at  home  by  just  analyzing those  kinds  of  data  and  skills  for  the  task. I  had  this  one  in  the  walking  slide. When  I  explain  what's  the  walking. Walking  is,  as  I  said  before, really  well  studied  from  different  fields. And  we  have,  these  are activated  muscles  and  we  have all  these  kinds  of  data  from  some  error. How  to  collect  this, how  to  solve,  how  to  know  the  skills  for  walking. It's,  I  think  currently  good  solution  is  data-driven  way. We  need  more  data  for  the  Fiona  sold  some  kind  of  task. And  you're  worried  experiments. Obviously,  you  just  study. So  this  is  one  scenario using  all  the  arms  for  a  different  test. We  can  attach  arm  to  the  word, or  this  is  just  a  Roomba. So  we  can  attach  one  arm  to  run  bar  to  make  it  as  a  base. And  this  experiment  is just  a  one  lunch  or  fully  autonomous. While  the  robot  is cleaning  the  table  and  doing  some  dishwasher. There. Three  times.  Faster  video,  but  still  slow. But  I  think  we  don't. It's  okay  to  take  some  time because  when  we  are  using  vacuum  machine that  we  don't  think  about the  actual  time  duration  for  the  cleaning. So  while  the  robot  is  working  hard, users  can  play  with  this  robot. So  what  I'm  showing  is  actually you're  lonely  life  in  the  future. So  it's  all  based  on  the  perception. So  you  can  move  things from  the  mobile  base  to  the  kitchen. And  I  think  there's  I know  there's  a  very  simple,  very  simple  one. You  don't  need  to  full  synopsis  them. You  can  say  set  actually  very  important  during  COVID, we  all  found  that  some  people who  actually  need  very  simple  physical help  our  couldn't  help  a  lot. So  if  we  can  make  this  kind  of  system actually  used  in  home  environment, we  can  help  pass  up  people. So  yeah,  this  is  the  task  retargeting. This  light  blue  ones are  the  current  topics  I  want  to  solve  it  first  as  a, much  more,  a  home  labor  related  works. Because  what  I'm  doing  is  configuration  free. And,  uh,  what  I  wanna  do  is  solving this  test  coin  a  bit  general  way. I'm  pretty  sure  we  can  use the  same  solution  in  the  humanoid  shape. My  goal  is  still  not  change  it. I  want  to  do  humanoid  robot, but  currently  is  way  too  expensive. And  I  think  we  can  do, we  can  tackle  this  kind of  problems  with  much  simpler  system  first, to  collect  more  data. And  we  can  use  that  kind  of solution  in  human-shaped  robot  later. So  that's  it  for  my  talk. I  can  get  questions. Yes. Yes. I  do  want  to  see  this. Now. Yeah,  we  made  that  actually  fly  in  my  lab. I  will  show  you  that  later. That's  really  good  question. I  also  thought  about, so  if  they  wanna  do  so, dishwasher,  kinda  robot,  robot, this  home  appliance  is  for  that  task. And  it's  a  much  more  efficient than  having  like  six-degree  of  freedom. Right? But  space  wise,  if  you  think  about  this  space, the  dishwasher  is  using  one  slot  in  the  kitchen,  right? So  if  we  can  have  really, really  large  house  than if  he  can  have  all  the  home  appliance. Some  error,  maybe  that's  better. Like  company  or  hotel  as having  like  cooking  robot  underground. It's  fine. But  if  you  wanna  do  that  at  home or  some  other  space  with  this  space  constraints, maybe  a  bit  dexterous  robots  which can  do  multiple  task  is  better. I  think  you,  today, when  you  think  about  bipeds and  planning  the  motion  of  the  body. Let's  step  planning  center  of  mass  inverse  kinematics. Different.  That  depends. So  if  e.g. the  Samsung  robot  didn't  have  the  self  step  planner. So  more  reactive  control or  slightly  pre-planned  motion. I  can  not  say  that's  pre-planned  motion, but  it's  like  making  the  robot  go into  inside  the  limit  cycle  is passive  dynamic  kind  of  approach,  right? But  if  you  want  to  make  the  robot  goal, ops  stairs,  or  there  are some  task  we  need  accurate,  like  footstep  planning. And  in  most  case, for  the  competition  or  a  challenge, we  need  that  part. So  if  we  want  to  do  this  for  the  totally  rough  terrain, like  space  or  some  error, I  might  just  use  energy  efficient  way, even  including  the  computation  side. But  for  now,  if  what  we're  seeing  is  better, stable  the  T,  right? So  I  think  a  planning  ways  currently,  the  best  solution. Shoulders. Stay  like  this  robot. Where  human,  wherever  this  mall  arm, more,  they  are  different  configuration. You  can  either  go  to  like  that  even  beyond  human,  right? That's  true. It  doesn't  matter  if  any,  Right? So,  yeah,  that's  also  a  really  good  question. If  we  want  to  use  the  robot  so  humanlike as  the  mechanism  is  a human-like  is  important  to  human  shape  is  important. But  if  we  want  to  use  the, this  kind  of  different  shape  or  configuration 1s,  we  don't  know. Users  usually  don't  know  how  this  one  works,  right? So  how  to  solve  this  inverse  problem? To  do  a  simple  task, we  sometimes  cannot  predict  the  motion because  the  motion  is  usually  using their  own  objective  function to  reduce  the  energy  or  something,  right? So  how  to  make  the  motion  human-like  for  the  users? To  predict  the  motion  for  the  users  safety. So  humanlike  is  likeness  is  a  four. Now,  with  this  system, I'm  more  focusing  on  the  user  point  of  view, how  to  make  the  motion  natural, or  how  to  provide  this  more  like personal  habit  kind  of  task. So  you  when  you  have  when  you  are  using  the  dishwasher, Everyone  is  actually  printing their  dishes  in  different  way. So  how  to  convey  that  kind  of  like  human  likeness  or  it, That's  my  current  interests. And  the  one  you  said  was, there's  no  mobility  here,  right? Well,  there's  a  small  mobility  here. But  human  actually  can  move  this  one  around. So  what  I'm  saying  is, if  we  use  this  one  and  if, let's  say  one  arm  is  like  $20,000, then  very  expensive  here, but  we  can  just  use one  num  in  turn  And the  human  can  just  move  around  disarm. Or  we  can  share  the  sum  between  households,  right? So  what  I'm  saying  is,  okay, let's  make  this  service  or  a  robot  available  in  many, how  sores  than  we  can  find  something. And  after  that,  if  if  Tesla or  some  other  company  make  sure  you  all  know is  a  weird  $20,000  than  that. We  can  use  that  solution  in  different  form. That's  kind  of  my  research  direction  now, could  you  say  again,  are the  number  of  users  or  too  small? That's,  I  think  that's  the  current  problem. Motors  are  expensive. So  frankly,  that  $20,000  per  arm  is  actually  in  my  lab. If  we  need  to  develop  these  one-by-one,  it's  a  $20,000. So  how  to  reduce  the  cost as  only  solution  is  mass  production. I  think  that's  not like  what  we  can  do  in  academia,  right? So  what  I  wanna  do  is  I  want  to  show, share  this  kind  of  idea  and  that  this is  simply  about  plugging. So  if  we  can  have  this  one, make  this  one  as  standard  to  use other  arms,  other  commercialized  dams. And  then  I  think  users  will, the  number  of  users  will  increase and  then  the  cost  will  go  down.