Let  me  introduce  our  prestigious  speaker, Professor  Hen  Park  from  his. Today,  he  will  give  us  a  talk  with  the  title  of the  hardware  Design  and  Control  algorithms for  Asia  and  First  Le  Robots. As  Tyler  suggested,  Hen  has  made numerous  impactful  work  in  L  Robot  bias  probotics, control  algorithms,  learning,  and  so  on. And  he's  also  a  recipient  of  the  NSF  Corrier  Award, as  well  as  the  RSS  Early  Corrier  Award. So  please  welcome  our  today's  speaker, Hen,  and  he  is  all  yours. Hello.  Can  you  hear  me? Yes,  we  hear  you. Okay,  thank  you  for nice  introduction  and  thank  you for  inviting  me  here  to  give. Okay.  Do  you  have  a  chance  to  talk with  talk  with  Tujia  Tech? So  it's  actually  my  start  time  to  visit  Atlanta. I  visited  here  for  during  my  poster  and then  today  Weador  is  actually  very  unusual  to  me. I  mean,  I  think  it's  also  unusual  to  you. But  yeah,  I  I  actually uh  little  bit  upset that  I  miss  chance  to  keep  the  seminar  in  person. But  anyway,  so  let  me  start. So  today,  my  topic  is  how  do  you design  and  control  algorithm  for  agile  and  psi  robots. So  I'm  basically  ego  researcher. So  I  do  control design  and  then  how  do  I  design  for  leg  robot. Still,  great  examples  of the  robot  system  is  actually  biological  animal, especially  as  you  can  see  in  this  video. Humans  have  extreme  balance  capability to  climb  the  wall  rock climbing  without  utilizing  his  hands. And  then  this  dynamic  balance  can  be  achieved  by this  funny  looking  very  fast  kicking performed  by  Ukrainian  dense  popomo. This  agile  and  then  very  dynamic  capability  provides a  great  balance  and then  very  agile  performance  in  biological  animal  system, including  human  and  then  some  quadruped  animals. But  this  video  shows another  level  of  accelerating mobility  and  then  this  is  a  squirrel. We  assume  that  this  animal  does  not have  a  very  intelligent  capability, but  this  animal  can quickly  jump  over  gaps  and  then  obstacles and  then  while  utilizing its  full  body  capability and  then  physical  limitation,  something  like  that. You  can  climb  the  lobe  and  then utilizing  as  you  can  see  here,  some  ladders. And  even  you  can  wide  on some  elevator  mechanisms  to navigate  through  very  complex  environment. So  here  we  can  see  that  very  intelligent  behavior. Can  combine  many  locomotion  skills to  come  up  with  navigation  on  very  complex  terrains. In  our  lab,  we  are  trying to  reproduce  the  capability  in  god  robot  system. Of  course,  we  do  how  do  you  design  and  then  manipulate implementation  and  then  so we  manufacture  our  own  hardware. Then  we  implemented  our  control  design  in  our  robot. This  video  shows  how  robot  Honda  actually  navigate through  climb  climb  of the  hilly  Mountain  nearby  the  Christ  Campus. So  in  this  talk,  I  will  show  two  robots. The  first  two  robot  is  stand, and  then  this  robot  is  designed  for  fast  running. Then  to  do  that, we  designed  our  own  custom  motors, and  then  we  installed  that  motor  to  our  robot. Then  second  robot  I  will  try  to  show  is  a  marble  robot, then  this  robot  is  designed  to climb  the  particle  surface  very  quickly. Then  this  robot  equips  with novel  magnetic  food  which  can  provide  the  audation  for. Then  to  implement  that, we  utilize  the  electropermant  magnet. Then  unlike  the  electric  magnet, this  magnetic  device  does  not  need electric  energy  to  keep  the  magnetic  force  on. So  we  only  need  magnetic  energy. We  only  need  electric  energy when  switching  between  of  state. You  can  save  some  energy  by utilizing  this  electropermant  magnet, this  switching  can  be  done  very fast  in  a  very  fast  manner. And  then  for  this  novel  hardware, we  design  our  own  control  golism. So  we  utilize  two  frameworks,  two  approaches. The  first  one  is  body  based  control  golism. Specifically,  we  utilize  the  body  printing  control. And  then  we  do  some  study  on  the  body itself  and  then  optimization  embolism  itself. Second  approach  is  actually  learning  based  abolism. So  we  employ  the  reinforcement  learning  albolism as  well  as  we  are  now  studying  about some  hybrided  type  bolism  which  can  combine product  optimization  and  reinforcement  learning  bolism which  is  ongoing  study  now. To  design  the  control  organism, we  have  to  come  up  with  a  control  algorithm  which  can respect  how  characteristic  as  well  as  the  limitations. Then  so  how  do  you  like  magnetic  food has  some  unique  function  which  did  not  exist. The  previous  hardwares. So  we  need  to  come  up  with  a  simulation  algorithm in  order  to  reflect  this  new  hardware and  then  how  do  you  charter  extend  limitation. So  in  this  way, we  can  actually  fully  utilize hardware  capability  over  existing  system,  okay? So  sometimes  we  have  hardware  specific  simulation  model. So  we  add  this  kind  of  simulation  model to  existing  software  for  dynamic  simulation. So  four  Egoien  per  actuator  dynamics  can  be  implemented. Okay.  So  actuator  is  not  just  a  pure  to  source. It's  actually  it  has  its  own  dynamics. Electric  motor  is  its  own  dynamics. So  we  add  that  in  the  simulation  model. As  well  as  double  hardware  model. We  mathematically  model  double  hardware and  then  add  it  in  the  simulator. Okay?  So  to  come  up  with this  particular  climbing  in  the  simulation  model,  okay. And  then  design  control  algorithm requires  Bd  or  control  design. So  in  the  design  of  control  algorithm, we  have  to  consider  we  have  to respect  this  specific  simulation  model. For  example,  your  function  has  to  be  carefully designed  so  that  it  can  reflect  how  specifics, as  well  as  constraint in  your  reinforced  multi  learning  algorithm, or  boda  based  control,  okay? Boder  based  control  design. So  today  I  will  talk  about  how to  implement  this  or  how  to  come  up  with  a  design, how  do  you  specific  simulation model  and  then  boder  or  control  design. Okay.  And  over  course, as  a  experimental  roboticist we  encounter  a  lot  of  momentum  catastrophic  failure, and  then  we  can  learn  something  from  this  experience. For  example,  this  is  mab  robot  for  my  PhD, and  then  we  broke  Mabel's  foot, Weber's  ankle,  don't  worry, it  takes  about  only  30  minutes  to  fix  this  ladder. Then  for  tattoo  robot, my  poster  has  some  amplifier  explosion. And  then  some  weird  behavior  that  was  not  designed  by  me. Then  this  is  t  to  bot  fell  because  of  the  pad  control. This  failure  provide  us  some  lessons, but  especially  these  two  failures actually  motivated  us  to  design  new  control  algorithms. We'll  get  back  to  these  two  failure  later. First  of  all,  let's  talk  about  the  audio  design. So  we  are  utilizing qua  duct  driv  who  are  so  called  the  proprestetutor, which  was  originally  proposed  by P  pressor  San  Be  Crowd  AIT. And  then  we  agree  that this  QDD  attor  was  actually  enable, technological  enabler  for the  success  of  quadrupedal  robot, as  well  as  some  human  aid  robot. Okay?  The  idea  is  actually  very  simple. We  utilize  high  torque  motor, as  well  as  Rugiar  ratio,  reduction  mechanism. In  order  to  improve  the  transmission  transparence, and  then  improve  the  vector  ability. So  here  I'm  showing  you the  properstive  actual  to design  with  Rugo  reduction  ratio. But  how  much  KO  reduction  ratio  we  have  to  utilize. In  this  question,  we  utilize  trajectory of  humanity  algoolism  so  in  the  previous  messod, the  design  procedure  is  actually very  iteratb  and  sequential. So  here  we  select  Q  ratio, and  then  the  selected  Q  ratio determine  the  operating  region  of  the  motor, with  respect  to  pork  and  then  Omega,  angular  velocity. We  did  this  photo  operating  region, you  solve  constraint  tractor  accumulation  algorithm and  then  obtain  the  trajectory. And  then  we  check  that  this  obtained trajectory  satisfy  your  desired  performance. And  then  if  it  doesn't  satisfy  the  design  performance, then  we  go  back  and  then  iteratively  select the  nuclear  ratio  and  then  do  this  again. Okay.  So  in  our  work, we  combine  all  of  this  process  in the  one  nolinearocen  problem, and  then  we  come  up  with  a  new  none  cogen  problem. Here,  we  try  to  maximize  jumping  height  while  reducing the  impactive  force  upon  landing  and then  respecting  limitation  on  actor  token  speed, then  we  formulated  the  unin  programming  problem. We  do  this  performance  specification. And  then  when  running  a  programming  algorithm, we  obtained  the  ratio  of  22.9  to  one, which  is  actually  a  little  bit  higher  than  MI  Theta  case. Which  was  5.8  to  one. Then  to  realize  this  QO  ratio, we  have  to  design  compound  the  planetary  mechanism. Then  this  is  the  performance  of  the  jumping  height  over the  layer  installed  with  the  tuy  to  be  designed. Then  this  layer  is  actually implemented  in  the  bubble  design,  okay? Next  design  optimization  we  did  is  actually now  we  try  to  include  RombobGyotis  in  our  non  problem. And  then  Dumbo  otis  is  integer  variable. So  now  the  problem  will become  big  integer  linear  program  problem,  right? So  you  have  mboicyots  of  the  lingear, sun  gear,  and  planic  gear. This  combination  will  provide  the  Keo  ratio, and  then  this  eo  ratio  will  affect the  many  things  like  bottle  selection, swing  retractory  energy  cost,  something  like  that. The  benefit  of  having this  dumbbicGots  in  your  octal  algorithm  is  actually  you can  consider  the  hardware specific  design  or  detailed  design. So  for  example,  mbobGots  of the  gear  be  rather  than  equal  to some  value  to  avoid  undercut, and  mboegyot  hips  and gear  should  be  larger  than  what  equal to  this  number  in  order  to  make  the  shaft  to  go  through. The  whole  picture  is  in  the  hip  gear,  for  example. So  this  kind  of  specific  design  consideration, can  be  included  in  opt  problem  if you  consider  Dumbo  Giots,  okay? And  then  finally,  play assembly  constraint  so  that  you  can  assemble  this umboyotGars  dumbiotsT  make  the  plant  set,  okay? And  then  if  you  solve  this  option  viabilism, you  can  obtain  the  TO  ratio  as  well  as  the  Dombobekotis. And  then  this  Dombois  can  be directly  used  to  design  the  pltric  your  set. Okay?  So  we  designed our  pltric  yourset  and  then  motor  module, and  then  we  install  that  in  the  Kast  hand  robot,  okay? So  this  is  how  we  design the  asi  director  drive  tretor  to  select  the  Oratio. Okay?  We  utilize  our  pltric  yourself. Since  I  don't  have  much  time, I  will  just  directly  go  to  the  control  design. The  first  control  method  is  a  body  based  control  design, specifically  body  pretty  control. In  body  control,  we  try  to  solve this  very  small  opt  problem, and  then  it  tells  real  functions,  Ta  transition  model. Then  if  you  solve  this  finite  time  open  problem, then  you  can  obtain  the  set  of  sequences  of  actions, and  then  you  implement  this  on  the  first  input, first  action  at  T  equals  zero. After  that,  at  the  next  simple  time, you  solve  again  this  optimal  problem  and  then  obtain the  sequence  of  actions  and the  only  apply  the  first  contraactions, and  then  go  so  on  and  so  on. Then  you  can  obtain  Btback  then this  open  problem  will  be  solved  within around  ten  millisecond  in  order  to  make  real  time. But  the  problem  is  the  transitent  model  is  actually  very complex  because  the  robot is  consist  of  multiple  sitive  body. It's  actually  high  degree  of  freedom  model. So  what  we  do  is  actually  reduce  body  complexity. So  we  took  out the  contact  dynes  model  and  then  we  assume  that the  robot  is  actually  just  one  sitive  body  and then  this  sited  body  is controlled  by  ground  direction  force, we  did  that,  we  actually  can  lead  contact  dynamics,  okay? So  because  we  don't  have  a  contact  dynamics, the  contact  sequence  should  be  predefined. It  cannot  come  from  the  optimal  agent  algorithm  because optimal  golim  doesn't  know the  contact  dynamics  does  not  have  a  contact  dynamics. So  we  normally  predefine predefine  the  sequence  of good  cogient  for  throating  gate, bounding  gate,  and  then  Calgon  gate. And  then  we  design  the  MPC  algorithm. Then  this  method  already  has  been  utilized to  come  up  with  autonomous  jumping  over  Mitiiato. Okay. So  if  you  look  at  the  robot, single  lit  of  dynamics, then  those  two  robots actually  have  the  same  mathemetical  structure, Chitund  and  then  bubble  robot, they  have  exactly  the  same  model, but  different  inertial  parameters. And  then  different  to  contact  constraint. Hound  robot  has  to  satisfy  fun  friction  constraint. Bubble  robot  has  to  satisfy foot  constraint  imposed  by  magnetic. It's  it  bit  different. So  but  they  share  the  same  dynamics, and  then  they  can  use  the  same  optima  algorithm. Okay?  So  here  in  single  letter  dynamics, if  you  are  working  on  the  three  D  single  it  dynamics, you  have  this  manifold  variable  R, which  lives  on  SO  three  manifold. And  then  normal  amps  ableism convert  that  into  ole  angles. Then  Olo  angle  some  problem, some  issues,  mathematical  issues. So  it  doesn't  provide a  consistent  distance  between  two  orientations, and  then  it  suffers  from  the  Kembala  problem. I  suffers  from  the  singularity, can  you  do  this  problem  in  just  normal  robot, it  will  not  be  problem because  it  works  on  the  flatground. I  does  not  go  like  porticle  surface  normal  robot. But  for  marvel  robot, it  has  to  climb  the  porticle  wall, then  it  will  go  through  the  poticle  posture, which  is  actually  singularity  for  oil  angries. This  experiment  actually Failed  because  of  the  problem. There  are  many  types  of  oil  angles, and  then  some  angles  has a  singularity  when  angle  becomes  90  degree. I  picked  up  wrong  oil  angle  choice. So  that  resulted  in  this  failure,  okay? Now  we  formulate  the  MPC  problem  on Asos  manifolder  Then  the  octgen  problem  becomes optimization  on  As  manifolder  then  we  utilize reparameter  trick  in  order  to  make this  octen  problem  to  oximogen  on  tacto  space  problem. To  solve  this  problem  efficiently, we  need  Jacobian  analytical  gradient  and Hasian  so  we  specifically use  custom  euthanasia  ego  lesion,  okay? And  then  we  actually  adopted  the  presented  in this  paper  published  in  Taro  in  2016. Now,  after  this, it  becomes  quadratic  programming  problem, and  then  we  utilize  our  own  custom  QP  robo. Okay.  And  this  is  the  result  of  running  first. We  do  three  meter  per  sec. And  the  robot  can  handle  very  wild  kicking. And  then  we  can  transfer this  control  algorithm  to  climbing  robot  as  well. The  only  thing  we  change is  actually  food  contact  constraint. So  we  derive  the  due  contact  constraint for  magnetic  foot  so that  it  doesn't  slide  or  peel  off,  okay? Then  due  context  constraint  is  actually intersection  of  these  two  constraints. Okay?  And  then  when  implemented  our  algorithm  to  robot, we  can  obtain  the  pace  gate with  a  speed  of  0.7  meter  per  se. As  well  as  inverted  throating  gait with  the  speed  of  0.5  meter  per  se, and  then  the  magnetic  foot  and  then  robot  is powerful  enough  to  handle  two  kilogram  payload for  particle  climbing  and three  kilogram  payload  for  inverted  throat  walking. Then  if  you  utilize  some  trajectory  opt  algorithm, then  you  can  come  up  with  this  trajectory which  can  traverse  the  gap. And  then  we  can  transition from  particle  world  to  embodied  world. The  robot  was  also  tested  on  the  outto  experiment. This  is  actually  water  tank  made  of  steel, but  it's  actually  covered  with  very  thick  paint. Paint  depth  thickness  is  actually  0.3  millimeters. But  the  magnetic  food  is  strong  enough  to provide  odation  force  still. And  this  method  can  be  also  applied  to state  estimation  because  stative  estimation  is basically  another  optimal  problem, then  we  proposed  the  invariant  smoother for  state  estimation  for  a  robot, and  then  you  provided best  result  compared  to  state  ibolism. Now  let's  go  back  to  the  second  failure. This  is  the  second  failure. What  happened  to  robot  is  actually  stepped onto  the  non  moving  part  of  the  treadmill. And  then  even  with  these  ericno  cases, the  robot  does  not  change  the  kit to  recover  from  the  cono  cases  because  it  does  not have  that  ability  because  we predefined  contact  sequence  pH  design,  right? So  now  we  want to  include  the  contact  model  in  our  NPH  problem. So  here  now  we  add contact  impulse  to  our  transition  model. And  then  this  contact  in  pulse  is obtained  by  solving  this  optimization  problem. This  opinion  problem  has  quality  cost, but  it  had  some  constraint. Then  we  call  complimentary  constraint. Because  of  this  special  structure of  complimentary  constraint, the  solution  of  the  problem  will be  categorized  into  three  cases. The  first  case  is  separating  case. After  the  impact,  the  robots  foot  leave  the  ground. This  is  easy.  And  then  clamping  case. After  the  impact,  robots  foot stuck  with  the  ground  and  doesn't  move. In  this  case,  after  the  impact,  city  becomes  zero. And  so  case  is  just  sliding  case, the  case  robot  slide  after  the  impact. So  for  each  cases, we  can  obtain  the  analytical solution  impact  contact  impasse, and  then  because  we  have  analytical  expression, we  can  obtain  analytical  gradient with  respect  to  some  physical  quantity. Okay.  So  for  clamping  case, we  can  see  that  after  the  impulse, the  peruse  becomes  zero, the  foot  ferocite  becomes  zero. So  any  gradient  contact  erosit with  respect  to  some  variable  will  become  zero,  right? And  this  result  in  local  minima,  actually. So  for  example,  let's  say  in  this  robot, the  reference  position  is  actually  very  up. The  robot  has  to  jump in  order  to  match  the  reference  position,  right? But  it  does  not  jump  because  it  is  in  clamping  model. I  started  with  the  colapping  model, then  gradient  die,  okay? The  desired  solution  is  something  like  this, periodic  jump  to  match  to  the  left  position. If  you  see  the  gradient, then  you  can  see  that  there  is  a  local  minima  on the  laptop  side  of  space. To  solve  this  problem, the  solution  of  a  complimentary  constraint, the  space  of  that  looks  like  this. It  can  be  either  potical  line  or  horizontal  line,  okay? We  actually  approximate  that  space  with more  smooth  and  continuous  space.  Something  like  this. Then  we  introduce  that. This  is  more  continuous  and  smooth  space by  introducing  the  variable  law, which  is  positive  value. If  law  is  very  ragy, then  it  actually  doesn't approximate  well  with  the  original  constraint. But  if  law  is  very  small, then  it  approximates  very  well  the  original  constraint. We  have  actually  adjusting  now, how  well  we  approximate  the  original  constraint. So  original  lectic  gradient  robot  tends  not  to  jump. And  there's  actually  previous  method  with  some  heuristis. This  does  not  provide  the  periodic  jumping  behavior. So  if  you  use  our  method, then  it  provide  very  periodic  jumping  and then  it  can  reduce  the  cost  a  lot. Okay?  And  then  if  you  look  at  this  gradient, you  can  see  that  when  oc  zero, this  is  actually  the  same  as  original  constraint. There  is  still  local  minima. On  the  left,  as  we  increase  the  low  value, it  actually  converted  it  to  global  minima, even  though  it  starts  with  close to  the  local  minimum  point. But  if  you  increase  too  much, then  the  solution,  the  quality of  solution  is  actually  become  worse. Okay.  So  there  should be  some  sweet  spot  over  the  values  of  low. If  a  pue  too  small,  it  does  not  jump. If  a  pelu  too  large, then  it  jumps  too  often. Result  in  bed  cost  value. There's  actually  pretty  wide  sweet  spot. The  probide  pretty  good  real  value,  okay,  cost  value. Now  we  integrate  this  analytical  gradient with  differential  dynamic  programming. Info  pass  we  have  exact  impasse  model. In  ***  pass  where  we  need  a  gradient, we  use  approximate  smooth  analytical  gradient. Then  we  employ  the  multi  pushing  variant of  DP  to  improve  the  cobotent  property. So  if  you  implement this  algorithm  to  robot  to  the  simulation  model, on  the  top,  you  have  a  reference  trajectory. There  is  no  swing  trajectory. There  is  no  food  contact  sequence, but  the  algorithm  automatically come  up  with  a  new  context  sequence. Here  slippery  surface, robot  sometimes  utilize  the  slippery  surface, sometimes  it  propel  so that  you  can  catch  up  with  the  reference  if  required. Then  this  cutely  motion the  algorithm  can  come  up  with  this  nice  cuty  motion. Then  we  extend  this  idea  into  the  three  D  model, we  add  some  additional  od  that can  prefer  protein,  symmetriate. That's  only  the  added  OD  functions. Then  you  come  up  with  this  nice  motion. Then  for  this  unstable  posture, this  is  actually  a  very  unstable  posture. The  robot  come  up  with these  nice  motions  to catch  up  with  this  unstable  posture. There  is  no  equilibrium  in  this  unstable  posture. The  only  way  to  follow this  reference  is  actually  making  contact. Then  you  can  obtain  humanoid  robot  motion. By  changing  the  aprons. Now  the  robot  only  utilize  its  hindfoot. We  implemented  this  algorithm  to  al  robot. Of  course,  we  did  a  tremendous  amount  of engineering  to  make  this  algorithm  in  al  robot. I  cannot  be  contained  in the  journal  publication  suddenly. But  anyway,  so  the  computation  time satisfy  the  real  time  constraint. And  this  unstable  pose  can  be  followed. There  is  no  equilibrium. There  is  no  dynamic  equilibrium. So  that's  why  the  robot  cannot  keep  this  posture, but  it  tries  as  much  as  possible to  follow  the  reference  tractoryRference  posture,  okay? So  second  approach,  we  are  currently  utilizing, it's  actually  reinforcement  learning. And  then  we  recently published  robotics  and  robotics and  automation  megagin  paper. So  my  student  really  wanted  you to  work  on  the  reinforcement  learning. And  then  this  is  his  first  controller. So  this  controller  can  come  up  with a  four  meter  per  sec  first  running. So  four  metapoet  is  already  better than  Buda  based  control,  right? Bode  based  control  speed timm  speed  was  three  meter  per  se. And  it  actually  was  able  to  handle  the  cono  cases, even  though  it  stepped  onto the  no  moving  part  of  the  treadmill, it  can  recover  from  these  co  cases. But  if  you  look  at  the  gait, it  looks  strange  to  me. It  looks  not  natural  to  me,  right? I  mean,  animal  utilize  this  kind  of  gait,  you  can  see. I'm  not  sure  why  they  are  utilizing  this  skate. It  doesn't  look  energy  efficient  to  me  and  natural  to  me, but  they  say  they  exhibit this  skate  when  the  animal  is  happy,  right? So,  but  we  don't  have a  happiness  view  the  function  in  our  design. So  to  evenly  distributed  to  to  the  symmetric  foot, we  add  in  gate  preference  in  the  other  function, and  then  we  can  come  up  with a  more  natural  gate  which  provide  the  faster  speed. So  the  speed  was  increased  to  5.16  meter  per  se.  Okay? We  have  1.16  meter  per  sec  speed  increase. To  further  increase  the  speed, we  actually  employ  the  motor  operating  constraint. So  in  the  LD  file, you  can  actually  specify angular  velocity  constraint  and then  boto  toque  constraint.  They  actually  separate. So  they  will  provide  a  box  constraint. But  we  are  operating  vision  of  motor  space, is  not  actually  box  constraint. It  looks  like  that,  because  of  Bagga  and  resistance. If  you  have  inductance  in  your  model, then  it  becomes  is,  Okay? But  we  only  consider  resistance  in our  model  and  then  we  obtain  this  operating  vision. So  if  you  see  this,  okay, token  angular  curb,  you  see  that  it does  not  utilize  much  of  the  region,  okay? So  we  implement  the  bottle  operating  region in  our  reinforcement  learning, and  we  gain  actually  higher  speed. We  obtained  the  0.5 meter/second  increase  speed  again,  okay? By  just  adding  that  constraint. Okay.  And  then  you  see  that it  more  fully  utilize  the  bottle  operating  vision. And  then  as  a  mechanical  engineer, we  have  a  choice  of  changing  our  mechanical  hardware, so  we  change  the  foot  so  that  it  becomes more  like  it  becomes  less  heavy. And  then  we  obtain  the  6.5  meter  per  set  learning. And  this  is  actually  two  year  back. So  we  broke  this  code  and  then some  other  institution  broke  our  code  as  well  recently. And  then  the  robot  finishes 100  meter  sects  within  9.887  seconds, and  then  it  made  out  robot  to  have  kinss  world  recorded. I  mean,  yeah. And  then  the  enforcement  running algorithm  come  up  with  a  galloping. So  if  you  have  a  dog  Wahed  in  your  home. You  know  that  they  don't  utilize trotting  gate  for  fast  motion,  right? So  we  can  actually  implement  the  gloping  motion, but  sadly  it  doesn't  approves  faster  performance. We  don't  know  why,  but  as  I  showed  you  already, we  utilize  approximator  algorithm  for this  robust  walking  on  the  mountain. Okay,  let's  skip.  Mm. A  so  we  are trying  to  implement  this  report running  algorithm  for  climbing  lob  as  well. If  you  use  non  magnetic  surface, then  if  you  use  MPC  algorithm, it  actually  fails,  o? As  you  can  see,  because  it cannot  attach  to  the  none  surface. But  in  the  eco  case, you  can  see  that  even  though  it  fails  to  attach  its  food, it  can  recover  from  it,  right? If  you  use  RL,  even though  it  failed  to  attach  to  the  surface, he  can  actually  recover  from  it,  okay? Then  if  you  integrated this  other  algorithm  with  vision  feedback, we  can  obtain  this  very  gain  performance. This  is  about  four  point  ometer  protect  learning, the  auto  experiment. Then  recently, many  humanoid  robot  has been  shown  by  many  companies  and  industries. We  don't  have  a  humanoid  robot  yet, but  we  can  show  that  just  utilizing the  actuated  technology  and then  reinforcement  learning  algorithm  that  we  have  now, we  can  probably  come  with the  humanoid  robot  control  borism. Even  though  we  don't  have  a  humanoid  robot, we  can  make  the  the  robot  work  like humanoid  without  changing  any  hardware  design. We  didn't  change  any  hardware  design, but  it  can  actually  learn  up  to  3.5  meter  percent  now. Then  you  can  recover  from picking  from  the  experimental.  Okay The  nice  thing  about  this  control  design  is actually  it  does  not have  enough  degree  of  freedom  for side  kicking  because  we  don't  have  a  actuator  for  that, but  you  can  actually  recover  from  this  kicking  as  well. Then  you  can  also  make the  robot  to  lift some  payload  which  is  about  7.5  kilograms. Okay. And  now  we  are  actually  trying  to  improve the  cod  quadrupeal  robot  running. Now  the  fastest  speed  we  could obtain  is  actually  9.5  meter/second. And  then  the  robot  can  finish the  hundred  meters  print  in  11.6  second. Which  is  probably  faster than  audiences  here,  including  me,  right? And  then  we  are  also  trying  to  break the  ecod  human  by  Hussein  Bolt. But  as  you  know, yeah,  we  have  a  lot  of  failure  moment, and  then  we  are  hoping  that  we  can  learn  from  something from  this  fail  moment  as  well,  okay? So  with  that,  I  want  to  conclude  my  talks, and  then  I  want  to  express my  uh  I'm  grateful  about  my  collaborators  as  well, including  SejonEojia  tech, and  then  thank  you  for  your  attention. I  think  we  have  time  for  a  couple  of  questions, maybe  via  Zoom  or  maybe  some  in  person  audience. We  have  a  handful  of  audience  here. So  any  questions? Yeah,  I,  wait,  wait. Let's  start  from,  let's start  from  the  in  person  question, and  then  we  will  go  back  to  Zoom. Right.  You  can  ask  Question. Okay.  Questions  about  have you  implemented  any  machine  learning  gology  to  come  up with  the  design  problem  combination  design  problem over  do  you  design  and  control  design,  right? Okay.  Not  yet.  Not  yet. I  mean,  we  have  not  done  yet. Okay?  And  then  the  difficult  thing about  how  do  your  design  work  is  actually, let's  say  we  come  up  with  some  value  and  then we  have  to  actually  manufacture  and then  assemble  the  real  hardware to  prove  its  efficiency  or  its  efficacy. Okay?  So  actually,  we  cannot  have many  cases  to  prove  this  work  this  method  works. So  it's  actually  very  difficult. And  then,  uh  I  mean, the  operation  algorithm  is  not  perfect. Sometimes  it  provide  okay  we  are  the  answer. So  I  have  not actually  extended  this  walk  yet.  Okay.  Yeah. Right.  I  think  there  was  a  Zoom  question. So  yeah,  maybe  we  can  start  from  the  top. Yes.  Okay,  so  the  first question  it's  actually  very  practical  issue. In  order  to,  uh, be  certified  by  Kennes  World  Record. You  need  resource  and  time  and  then  effort. They  are  not  very  dispensble. Then  you  need  some  money  actually,  like  publishing  paper. So  if  you  are  planning to  break the  Kines  World  recorder,  you  guys  can  contact  me. Okay.  We  actually  consulted  have  a  consultation  with the  Jonas  anos  group  from  Oregon  State  University because  they  broke  the  World  Lecod of  humanoid  running,  right? So  we  actually  we  actually  took  advice  from  them, and  then  we  successfully  contacted  the  Kyne  Word  Le  Code. So  it's  actually  a  practical  problem. And  then  even  though  you  break  the  ecode  once  again, I  don't  think  it's  not  related  to the  publishing  another  academic  journal  papers,  right? So  that's  the  reason. Right  next  is,  I  think  Frank. Is  it?  I  think  so.  Frank.  Hey,  hope  you  can  hear  me. Yes.  Awesome  talk,  really  awesome  talk. My  question  is  you  switched  from MPC  to  reinforcement  learning. And  so  I  wonder  whether  you  were  able  to  learn something  from  reinforcement  learning  that  we can  use  in  MPC  world. So  what  is  your  lesson  from  switching  to  RL? Yeah. So  I  mean,  theoretically,  I'm  not  sure. I  can  learn  from  this  reinforce  running  algorithm. But  for  practical  reasons, I  think  in  the  world  of  NPC  WA  based  algorithm, I  think  we  need  some  way  package  that  open  source. I  think  suit  of  code, I  think,  to  make  them  more  widely  used. Okay?  So  in  the  area  of  reinforcement  learning, they  have  really  good  package, really  good  open  source  code  base you  can  download  and  then  implement  it. I  think  that  can  actually  spread  they  can  actually  spread the  ages  to  more  wider  audiences  or  more  wider  engineers. I  think  we  need  that  for  Bod  based  control  society  more. And  then  RA  algorithm, I  think  is  really  good  at  come  up  with some  emergency  behavior  that  was  not  expected  by  humans. So  actually,  when  the  robot  collide  with  stairs, it  actually  changes  the  swing  gate automatically  so  that  it actually  picks  up  the  leg  higher  than usual  and  then  to  step  onto  the  next  stairs. Okay?  So  this  algorithm this  kind  of  behavior  was  only  enabled by  additional  design  of a  swing  tractor  swing  tractor  swing  trajectories,  right? But  in  RL,  you  don't  need to  have  explicit  algorithm  for  this. It's  actually  embedded  in your  urotwork  like  weight  functions. So  this  was  not  expected  by, B.  I  hope  this  answers  your  question. Yeah,  that's  awesome.  Thank  you. Right.  Yeah,  let's  do one  more  Zoom  question  and  we  will  get  back  to  you. One  mechanism  is  keeping  the quadruped  attached  to  body  c  and  inverted  surfaces. So  you're  asking  about  the  marble  robot,  right? So  we  use  magnetic  food. We  use  magnet, then  the  surface  is  actually  consists  of  steel. So  the  robot  only is  able  to  work  on  magnetic  surface  steel  surface. Okay?  So  I  mean, people  like  they  have actually  complained  when  they  see  our  PDO. Like,  why  don't  you  robot  work  on  more  piative  surface, like  concrete  surface,  something  like  that. But  we  believe  that  there's  actually many  industry  application  this  magnetic  worker because  in  many  industry  environment, for  example,  shipbuilding  industry, the  ship  is  actually  made  of  steel, but  they  don't  have  any  they  actually still  utilize  the  scaffold  to, you  know,  people  work  on  the  ship  surfaces  is actually  very  dangerous  and  then  very  costly. So  we  are  hoping  that  this  kind  of  robot  can  be  used  in the  shipbuilding  industry  like factory  inspection  consisting  of a  very  rgy  storage  tanker,  something  like  that. So  Let's  go  back to  one  in  person  question  and  we  will get  back  to  another  online  question. Can  you  ask  us  question?  That's  true. That's  actually  very  difficult  to  answer. I  mean,  I'm  not  sure  there  exist any  optimal  length  of the  lager  to  provide  optimal  behavior. It's  a  difficult  question  because the  design  parameter  for Length  of  the  leg  is  actually multiplied  to  the  control  toque, it's  actually  very  complex  bilinear  mathematical  problem. So  I  don't  think  there  is actually  true  optimal  for  this  problem. We  have  not.  Thank  you. Thank  you  for  the  to, can  you  please  reiterate  how  you overcome  your  problem  with  symbol  luck? Thanks.  Okay,  so  uh  the  previous  obtain  problem, they  reformulate  utilizing  the  oil  langers. We  don't  utilize  oil  lingers. We  formulate  all  the  cost  function  and  then constraint  only  utilizing  rotation  metrics. Then  this  rotation  matrix  representation  can  be converted  into  the  vector  space  representation, utilizing  reparameterization. Okay.  So  that  can make  the  problem  become  vector  space  optimization. Now  you  can  utilize traditional  optimigen  soluble  to solve  this  problem,  okay? So  yeah,  that's  how  we  avoid  the  Symbola  problem. The  next  question  is  when  the  robot  is  at  its  max  speed. Is  there  any  control  policy  that  enable  you  to  stap safely  and  rapidly  without  breaking  the  hardware? Okay.  I  think  currently, the  robot  can  accelerate  and decelerate  in  a  pretty  quick  manner. We  haven't  measured  the  maximum  acceleration, maximum  acceleration  to  provide  a  stable  transition. But  now,  it's  pretty  good. I  just  satisfy  our  needs. But  I  think  another  way  to provide  better  dilation  acceleration would  be  designing  transition, contrary  policy,  which  provide better  transition,  quick  transition. But  Actually,  we  need  to  move  soon. I  would  like  to  actually,  sorry, I  believe  you  have  many  questions, but  I  believe  the  ejs  question  will  be  the  last  one. Any  thoughts  on  other  augmented  NPC? Yeah.  Hey,  Helen,  this  is  a  very  nice  talk. So  I'm  just  curious. Since  you  mentioned,  the  RO  MPC  model  based  approach is  really  good  at  giving  very  reliable  motions, but  maybe  less,  you  know, versatile  or  emergent  as  what  RO  can  do. And  I'm  just  curious  if  you  have  any  thoughts,  you  know, combining  them  or  maybe or  maybe  for  certain  types  of  task, you  just  use  RO  for  certain  types  of  tasks, you  just  use  MPC. Or  you  think  maybe  combined  ROVs  and  PC  can  solve, you  know,  a  lot  of  different  kind  of  task. Okay,  yeah,  thank  you  for  your  question. So,  uh,  I  mean, we  have  some  preliminary  work collaborated  with  the  seventh  group, as  you  know,  we  have  imitation  learning  of  MPC  and then  we  train  further that  policy  with  reinforcement  learning, and  then  we  are  still  working  on  that  direction. And  then  another  direction  we  are  taking  actually  kind  of utilizing  auto  encoder  structure,  like  foundation  model. So  if  we  have  lots  of  trajectory  data  samples, and  then  we  can  learn  from  autoencoder,  okay? And  then  to  imitate  it  to actually  sample  the  data  samples, and  then  we  can  probably  add reinforcement  running  fine  tuning of  the  larger  language  model. Probably  we  can  come  up  with  some  new  controller. Actually,  the  vision  feedback  running  that  I showed  in  this  top  is actually  controlled  by  the  kind  of  algorithm  now. But  we  haven't  published  any  paper  yet. We  are  working  on  writing  paper. Right.  Thank  you  for  all  the  questions. And  also,  let's  thank  the  speaker again.  And  yeah,  there's  one. Yeah.  If  you  have  more  questions? Yeah,  sure,  feel  free  to  reach  out  to me  or  the  directly  Professor  Heron  Park. And  thanks  for  all  the  participation and  see  you.  Alright.  Thank  you.