You  hear  me  now. We're  good.  Thank  you. Okay. Hello.  Hi.  Okay.  I  think  we're  going  to  start. So,  first  off,  apologies. This  projector  is  not  working. We  tried  our  best,  but  it  doesn't  seem  to  be  working. So  everyone  sitting  over  here, if  you  wanted  to  try  to  move  your  chairs  over, that  might  be  good  cause  we  just  have  this  projector. Or  you  can  just  listen  to  Bobby. There's  a  few  chairs  in  the  back  over  there,  too. Yeah.  But  you're  gonna  have  a  hard  time  seeing. Yeah.  You  want  this  screen? Do  you  care?  I  don't  think  there. Oh,  you  just  want  it  lower. Can  stutter. I  don't  need  it.  'Cause  I have  I  have  my  own  preview  here. Yeah,  I  don't  know  if  There's some  kinematic  singularities. Yeah.  In  the  table  move? It's  side  down. Oh,  it's  rigidly  attached. Alright.  This  is  just  my  warning. Move  now  or  forever  hold  your  peace. Okay.  So,  introduction. So  today,  we  have  Professor  Robert  Greg visiting  us  from  University  of  Michigan. For  anybody  who  doesn't  know  him, I'm  gonna  do  kind  of  a  short  little  introduction. But  I  assume  everyone  knows  him  'cause  he's  great. So  yeah. Professor  Robert  Gregg  has  been the  Associate  Director  of  robotics  at  Michigan. You  became  it  in  the  fall  of  2020, which  was  before  the  Department  of  Robotics  was  founded. So  he  was  one  of  the  original  founding  members of  the  Department  of  Robotics at  the  University  of  Michigan, which  was  founded  in  July  of  2022. Um,  so  prior  to  joining  Michigan, he  was  my  notes  just  went  away. He  was  an  assistant  professor  at the  University  of  Texas  Dallas, and  then  he  received  his  BS  degree  in  ECS, which  is  electrical  engineering and  computer  science  from  UC  Berkeley, and  then  his  master's  and  PhD  degrees  in  ECE  from  UIUC. So  we  were  just  talking.  He's  very  multidisciplinary, has  many  different  departments that  he's  been  associated  with, which  is  kind  of,  I  think,  more  robotics,  right? So  his  research  is  centered  around control  of  lower  limb  robotic  assistive  devices, such  as  powered  prostheses. I  think  we're  going  to  see  a  lot of  powered  prostheses  today, but  it's  also  been  extended  to just  autonomous  bipolar  robotics. Research  is  conducted  through  his  lab. So  if  you  want  more  information, his  lab  is  named  the  Locomotor  Control  Systems  Laboratory at  University  of  Michigan. His  work  has  been  recognized  by several  awards,  so  I'm  just  going  to  name  two, but  the  NSF  Career  Award  and the  NIH  director's  New  Investigator  Award. Those  are  both  pretty  big  awards  to  receive. And  his  research  has  been  funded  by several  large  NIH  ROI  grants, which  are  very  good  when you're  working  with  patients  to  have  that  kind  of  fund. So,  thank  you  again  to  Professor Greg  for  visiting  us  today. We  look  forward  to  hearing  more  about  your  research. Thank  you,  everyone.  Thank  you,  Megan, for  the  invitation  and  the  kind  introduction,  as  well. It's  great  to  be  here  at  Georgia  Tech  my  first time  and  looking  forward to  talking  to  some  of  you  afterwards  as  well. So  yeah,  I'm  going  to  just  jump  right  into  it. So  to  start  with  some  motivation  for  the  talk, I  want  to  first  discuss challenges  with  amputee  locomotion. So  using  conventional  devices which  are  mechanically  passive, this  patient  population  tends  to  have difficulty  with  variable  cadence  walking, navigating  ramps  and  stairs,  standing  out  of  a  chair, and  because the  conventional  devices  are  mechanically  passive, they  have  to  compensate  with  their  intact  side  or their  intact  joints  to  make up  for  the  deficit  and  mechanical power  from  the  prosthetic  limb. And  this  overuse  of  their  intact  joints increases  the  risk  of  osteoarthritis  and  back  pain. So  powered  legs,  on  the  other  hand, have  the  potential  to  restore normative  biomechanics  by  being able  to  provide  positive  mechanical  work, active  control,  but  they  add  substantial  complexity. You  have  to  measure  things.  You  have  to  control  things. There's  compute  in  the  middle. And  so  that  has  significantly  limited the  deployment  of  powered  prosthetic  legs commercially  to  date. And  especially  challenging  when  you  have two  powered  joints  like the  knee  and  the  ankle  shown  in  this  picture, which  is  the  device  actually that  Aaron  Young  and  I  worked with  back  in  the  RAC  days,  which  was  fun. And  so  I'm  going  to  talk  a little  bit  now  about  how  we  tend to  control  these  complex  devices  and  um, typically,  there's  a  hierarchical  control  architecture, high,  middle  level  control  approaches. At  the  high  level,  it  tends  to  deal  with the  intent  or  the  activity  of  the  user. So  you  might  need  to  differentiate  between  slow, natural  and  fast  walking, but  then  there's  also  inclinations in  the  environment  that  you  have  to  differentiate. There's  also  different  types  of  inclinations  like  stairs, and  then  there's  sitting  and  standing  as  well. And  so  typically,  you  have a  different  control  mode  or policy  for  every  one  of  these  possible  activities, which  is  already  starting out  with  a  reasonable  amount  of  complexity. And  then  for  each  one  of  those  modes, you  have  a  mid  level  controller  which  has its  own  discretization  of that  activity  in  two  different  phases  of  gait. And  so  for  example, you  might  see  this  finite  state  machine  here, which  articulates  the  five  common  phases  of  gait, where  each  one  of  those  may  have their  own  set  of  control  parameters  like impedance  parameters  for  compliant  interaction with  the  ground  or  position, sorry,  or  proportional  derivative parameters  for  swing  phase. And  each  one  of  those  has  their  own  set  of parameters  that  need  to  be  tuned  to  an  individual. And  then  every  activity has  their  own  finite  state  machine, which  again,  has  its  own  set  of  parameters. And  so  already,  we've  now  seen this  explosion  in  the  parameter  space  of  these  devices. You're  talking,  you  know,  over  100  distinct  parameters. Not  all  of  them  maybe  need  to  be  tuned  for  every  user, but  a  significant  number  of  them  do. And  so  in  past  papers, it's  been  reported  that,  you  know, a  well  oiled  team of  researchers  and  clinicians  may take  5  hours  tuning  these  things. And  that's  clearly  not  a  clinically  viable  option,  right? There's  also  some  challenges  with the  discretization  of  gait  that  makes  it  less  smooth, compared  to,  you  know,  natural, smooth,  continuous  locomotion. And  so  my  research  has  been focused  over  the  past  decade  or  more  on  trying to  go  from  this  discrete  characterization of  human  locomotion  to more  continuous  representation  and  allowing  us to  control  amputee  gait  in  a  more  continuous  manner. And  so  we  started  this  by  looking  at a  continuous  representation  of  the  gait  cycle  phase. So  again,  that's  that  mid  level  I  was  talking  about where  we  have  the  gait  cycle. And  so  instead  of  having  a  discrete  phase  as  a  gate, we  have  a  continuous  representation that  is  measured  in  real  time. And  one  way  we  do  this  is  through  measuring the  residual  thigh  angle  on  the  amputated  side. And  so  you  can  imagine  that that  thigh  angle  is  reflecting  hip  motion, which  is,  you  know, under  the  volition  of  the  user. And  so  that  acts  like  the  crank  to  the  cycle,  right? The  hip  motion  is  swinging  forward  and  back. And  so  it's  like  a  pendulum  or  an  oscillator. And  you  app  that  to  a  phase  clock  that  we  see  here. I  guess  you  can't  see  my  mouse  on  that  screen. So,  I  apologize.  So  that  allowed  us  to  unify the  gate  cycle  under one  continuous  nonlinear  control  approach. And  then  more  recently, we've  been  working  on  trying  to  have continuous  representations  of  the  activity as  well.  So  at  the  higher  level. And  so  you  can  imagine  that  for any  estimated  value  of the  ground  inclination  or  the  walking  speed, it  would  be  nice  if  you  could  map that  estimate,  which  is,  you  know, a  real  number  or  real  numbers  for  these  two  variables  to, you  know,  the  biomechanically  correct  joint  patterns. That  might  be  joint  angle  trajectories. It  could  be  impedance  values, which  we'll  talk  about  later  as  well. But  the  idea  is  that  we  have  this  activity  space, which  is  a  multidimensional  space, we  want  to  be  able  to  model  how  the  joint  patterns  adapt as  the  activity  varies within  this  multidimensional  activity  space. Doctor  Greg,  I'm  sorry  to interrupt,  but  your  Zoom  slides? What's  our  First  talk  about  the  sense  of  phase. So  I  had  already  said  that, This  is  based  on  the  residual  thigh  angle, and  we  can  measure  that  with  an  IMU that's  attached  at  the  top  of the  prosthetic  knee  because  that's then  connected  to  the  socket  of  the  amputated  limb, and  it's  not  a  rigid  connection, but  we  assume  it's  rigid  and  it's  close  enough. Um,  and  so  you see  we  have  a  descending  region  where  we  can map  that  to  the  first  50%  to  60%  of  the  gate  cycle, and  then  you  have  sorry, first  40%  of  the  gate  cycle, then  you  have  a  kind  of non  monotonic  region  where  things  are  less  clear. So  then  we  use  feed  forward  or  time  based  mapping  there. And  then  we  have  an  ascending  mapping,  and  then  again, another  feed  forward  period where  we  are  using  time  because  again, we  lose  monotonicity  and  so  we uniqueness  of  the  mapping  from  thi  angle  to  phase. And  so  in  the  end,  we've  mapped this  sinusoidal  thigh  trajectory to  this  fairly  linear  phase  estimate, which  should  be  linear  for  steady  state  walking. And  it's  robust  to  multiple  ground  inclinations. It's  stuck  on  Zoom.  What's  shared. Oh,  shoot.  Okay. Zoom  are  stuck  on  it's  fine. Did  it  just  happen  or  started  that  way? Let's  try  that.  Better. Good  to  go.  All  right. So  yeah,  so  we're  able  to get  a  reasonably  consistent  sense  of  phase. And  this  is  also  nice  because  then  if  people  slow  down, phase  will  slow  down  the  progression through  the  joint  patterns  will  slow  down, and  if  they  speed  up,  the  progression through  the  joint  patterns  will  speed  up. And  you  can  even  do  some  things  like  stepping  backwards, which  we'll  see  later  and  so  on. Stuck  for  me  now,  yeah. Alright.  Hopefully,  there.  Yes.  Okay,  we're  back. Okay.  All  right,  so  I would  love  to  omize  this  thing,  too.  See. Hide. Nice.  All  right.  We'll  get  through  this  together,  okay? So  technology,  right? So  he  is  how  we  use  the  phase  variable. So  during  Sans  phase, we  have  an  impedance  controller, which  is  emulating  the  behaviors  of  a  spring  and  damper. But  these  are  variable  springs  and variable  dampers  and  also  a  variable  equilibrium  angle. So  this  is  as  called  variable  impedance  control, which  allows  compliant  ground  interaction, which  is  important  for  comfort  and also  robustness  to  uncertainty  with  the  ground. And  so  we  parameterize  we're  going  to  parameterize models  for  these  parameters  as  function  of  phase. During  swing  phase,  I  don't  know. There  you  go.  During  the  swing  phase, we  have  position  control, which  is  useful  because it  gives  predictable  swing  foot  positioning. You  want  to  make  sure  that  the prosthetic  foot  is  ready  to accept  your  body  weight at  the  conclusion  of  the  swing  phase. So  having  this  consistent  mapping  from  hip  motion  to  foot position  is  desirable  for  this  application. And  so  here  the  desired  angular  position of  the  joints  is parameterized  by  this  continuous  phase  variable. And  then  we  use  constant  proportional  damping  gains. And  so  together,  we  call this  hybrid  kinematic  impedance  control  or  HKIC, which  is  a  term  I'll  be  referring  to  later  in  the  talk, so  you  might  want  to  keep  that  acronym  in  your  mind. So  in  terms  of  generating  these  gait  models, we  take  a  data  driven  approach. We  have  a  parameter  we  have  a  parameterization  that's based  on  mathematical  formula, but  we  use  data  to  train  the  parameters. And  so,  for  example,  we  take  a  large  dataset, like  here  as  an  example, we  have  walking  over  multiple  walking  speeds and  ground  inclines. And  with  that  data,  we're  able  to  generate this  parametric  gait  model using  convex  optimization  techniques. And  once  we've  generated  that  model,  that's  all  offline, then  online  in  real  time, we  estimate  the  gait  phase  and  the  task  variation. So  when  I  say  task  variation, I  mean,  what  is  the  walking  speed? What  is  the  ground  income? And  then  the  gait  model  then  will  spit  out  kinematics  or impedance  profiles  for  the  prosthetic  leg  to  follow. And  so  this  is  one  example  of  that. So  we  see  here  we  generate  these  surfaces. And  in  fact,  these  are  higher  dimensional  than  three  D, but  I  can  only  show  you  a  three  D  plot,  unfortunately. So  what  we've  done  is  we've  fixed the  walking  speed  for  this  example, but  there's  also  a  walking  speed  dimension. And  so  for  any,  you  know, measurement  of  phase  or  task  variable, we  have  the  generate  we  can  generate the  biologically  inspired  reference for  impedance  parameters. And  swing  phase,  we  do  the  same  thing, but  for  joint  angular  positions.  Okay. All  right.  And  so  when  we  implement  this  in  real  time, we  have  ways  of  estimating walking  speed  and  ground  slope using  inertial  measurement  units. And  so  that  allows  us  to  have  this  continuously adapting  walking  gait  as the  treadmill  speed  or  incline  changes  in  this  video, it's  kind  of  a  slow  change, but  trust  me,  it  is  changing. It's  as  fast  as  the  him  will  incline. The  device  is  able  to  continuously  adapt  to  that. And  so  it's  able  to  accommodate within  some  reasonable  bounds, it's  able  to  accommodate any  combination  of  walking  speed  and ground  slope  any  biobmtic  way. And  so  in  the  end, this  is  showing  how  the  prosthetic the  recorded  prosthetic  joint  angles  and  moments  torques compare  to  the  able  bodied  references over  several  ground  inclines. And  so  these  are  actually  measured  from the  prosthetic  legs  and  coders and  the  torque  we're  commanding,  so  we  know  what  that  is. And  this  is  just  the  offline  data that  we  use  to  train  the  system. And  you  can  see  that  we're  able to  reproduce  these  patterns. It's  not  perfect,  of  course, but  the  patterns  hold  that  we see  flexion  increasing  or  decreasing as  appropriate  with  the  change  in  incline  and  we  see similar  patterns  with  the  joint  torques  as  well. Yes.  References  here. This  entire  row, this  entire  rows  reference  this  entire  row  prosthetic. Then  the  colors  represent  different  inclines. Yep.  And  so  as  a  byproduct of  restoring the  somewhat  normative  joint  kinematics  and  kinetics, if  we  look  at  joint  work, we're  able  to  show  that  we're  able  to,  you  know, either  decrease  joint  work  as  incline  decreases or  increase  joint  work  as  incline  increases. And  so  this  is  the  total  work  of  the  two  joints, in  particular,  we  see  that,  you  know, the  knee  should  have  net  negative  work and  the  ankle  should  have  net  positive  work, and  we  see  that. And  then  as  a  comparison  point, AB  is  the  able  bodied  reference. HKIC  is  the  controller  I've  been  talking  about, and  FSMC  is  just  a  finite  state  machine  controller that  we  hand  tuned  as like  a  baseline  reference  controller. You  know,  it's  our  best  attempt at  a  state  of  the  art  controller, again,  it's  not  to  say  that others  couldn't  do  a  better  job  tuning  these  things, but  it  was  our  best  attempt,  right?  So,  and But  in  some  cases, there's  not  a  lot  of  difference between  two  controllers,  right? Like,  right  here,  there's  not  a  lot  of  difference. There's  more  of  a  difference  at  the  ankle, and  then  the  total,  of  course. But  something  to  keep  in  mind  is  that  there  was no  tuning  for  the  HKIC  controller. It  was  all  data  driven  with the  able  bodied  references.  There  was  no  tuning  at  all. Whereas  the  FSMC  control required  probably  30,  40  minutes  of  tuning. Okay.  So  we've  also  been able  to  do  continuous  variations of  stericent  and  descent. Using  similar  kinematic  models for  swing  phase  and  impedance  models  for  stance  phase, where  instead  of,  you  know, a  variable  ramp  angle, we  have  a  staircase  angle  or  step  height. These  are  related  things. And  so  as  long  as  you  were  able  to  estimate  what that  step  height  or  stair  angle  is, then  again,  we  can update  the  patterns  of  the  prosthetic  leg. And  so  we're  working  on  ways to  do  that  with  dead  reckoning  with IMUs  and  also  some  perception  methods  as  well. And  then  for  sitting  and  standing, we  can  also  accommodate  continuous  variations. Here,  the  variations  would account  for  different  height  chairs. Okay?  So  some  chairs  are  taller  than  others,  right? And  do  we  don't actually  explicitly parameterize  the  chair  height  in  the  model. What  we  do  have  is  we  have a  slight  change  between  standing  up and  sitting  down  in  terms  of the  equilibrium  angle  we  see  here. So  we  have  slight  change  here, which  makes  it  a  little  bit  easier  to  initiate the  sitting  motion  because  you don't  want  to  have  too  much  resistance  to  sitting. But  the  thing  that  really  allows the  different  chair  heights  is the  way  that  we  automatically normalize  this  phase  variable  is  we're  able  to tell  what  the  starting  angle  of  the  thigh, you  know,  is  for  one  chair  versus  another. And  then  that  could  help  us  normalize the  thigh  ranger  motion  to  a  phase, a  variable  between  0%  and  100%  or  zero  and  one, I  guess,  is  our  convention  here  for  phase. And  so  here's  a  demonstration of  tall  medium  and  short  chairs  for  apitee  participant. And  here  we're  not  pre  programming  it. It's  all  normalized  automatically  to  the  chair  height. And  I'm  going  to  talk  a  little  bit about  outcomes  later  with  a  different  study, but  just  as  a  little  hint  at  that, we're  able  to  increase  the  speed  at which  the  participants  are  able  to  do  sit stand  repetitions  and  we're  also  able  to  reduce  asymmetry between  the  sound  side  and  the  prosthetic  side with  the  help  of  the  powered  device. I'll  show  you  more  data  on  that  later. And  so  taking  all  of  these, you  know,  kind  of  adaptable  mid  level  controllers. So  essentially,  what  we've  been  doing is  I've  been  describing these  mid  level  controllers  that  are  adaptable  to different  variations  in  the  activity, call  these  continuous  task  variations. This  allows  us  to  actually  then  reduce  the  set  of activity  modes  required  for  classification. So  we  go  from,  um  walking,  you  know, which  previously  had  several  different  modes to  accommodate  different  speeds  and  ramp  angles. Now  we  have  one  walking  controller that  accounts  for continuous  variations  in  incline  and  speed. And  then  we  went  from  having,  you  know, usually  there's  distinct  sitting  and  standing  modes in  a  high  level  controller. We  were  able  to  unify  that  as  well. With  stairs,  technically,  we're  able  to generate  a  single  control  model  for stair  ascent  and  decent,  much  like  we  did  with  walking. But  you  really  want  to  separate those  for  classification  purposes because  it's  too  late  by  the  time you've  hit  the  ground  with  stairs,  right? You  want  to  you  don't  want  to  find  out  that  you're  on a  staircase  after  you've hit  the  staircase.  It's  way  too  late. So  we  do  actually  still  separate these  in  terms  of  our  classification  space, but  they  are  technically  the  same  model.  Alright. And  so  then  how  do  we  do classification  between  these  activity  modes? Well,  because  we've  been  able  to reduce  this  classification  space  to  just  four  modes, we've  reduced  a  lot  of  complexity. There's  no  need  to  differentiate between  ramps  and  walking  on  level  ground  anymore, which  is  a  difficult  thing  to  do. It's  a  common  source  of  air  and  classifiers. It's  much  easier  to  differentiate  between  walking  and stare  ascent  or  stare  descent,  same  with  sit  to  stand. And  in  fact,  what  we  do  is  we  use heuristic  rule  based  class Heuristic  rule  based  transitions, which  are  kind  of,  you  know, hand  design  through  biomechanical  intuition. But  these  rules  are  very  simple that  we  can  explain  them  to  our  participants. They  can  learn  them,  and  then  they  can  then correct  errors  based  on their  understanding  of  these  heuristic  rules. Another  thing  in  our  toolbox  that's  really  important here  is  some  minimal  perception  of  the  environment. So  especially  for  stericent  transitions, it's  in  order  to  have  sufficiently  early  transition, we  find  it  very  helpful  to  have  some  perception  of the  environment  through  an ultrasonic  sensor  that  we  see  here. This  is  a  really  tiny  little  thing. It's,  you  know,  milligrams  and  weights, milliwatts  and  power  consumption and  just  a  few  dollars  cost. The  same  sensors  that  are  in  our  cars  these  days,  right? And  so  this  gives  us  the  distance  to  the  nearest  object. And  if  we  can  consistently  see  that the  nearest  object  distance  is  decreasing  over  time, that  means  you're  approaching  that  object,  right? And  so  that  gives  us  that  coupled  with kinematic  rules  tells  us when  we  have  high  confidence  that there  is  a  so  cent  transition  happening. And  so  we're  able  to  then  actually  have a  sufficiently  early  transition  stride  to then  you  get  proper  knee  flexion to  clear  the  step  and  so  on. And  also,  um,  this  allows  us  to do  what  we  call  ambilateral  activity  transition. So  either  the  prosthetic  leg or  the  intact  leg  can  lead  the  transition. They're  not  forced  to  lead  with  one  specific  leg, which  is  common  in  commercial  products. They  usually  have  to  initiate  the  motion  with a  specific  leg  using  a  specific  type  of  motion. And  so  uh,  you  know, I'll  quickly  summarize  some  of  our  findings  with a  recent  with  a  recent  paper  coming  out, recently  accepted  paper  at  transactions  on  Robotics, that  we're  able  to  get  about  99%  accuracy for  for  the  combination of  steady  state  and  transitional  strides. And  we  have  the  recovery  rate  is  100%. And  what  I  mean  by  that  is  we  have  backup  logic. So  this  99%  number  is  for  early  detection  of  transitions. So  an  ideal  detection  where  we  can  change the  mode  at  the  perfect  timing  to make  those  fluid  natural  transition. But  you  can  also  do  a  late  transition, which  is  just  not  as  nice. And  so  we  have  backup  logic,  also  heuristic  based, and  that  allows  about  50%  of misclassifications  or  maybe  higher  to  be  corrected. And  then  we  also  have  a  manual  override sentially  tell  the  participants  to  abduct  their  hips, and  it  will  reset  the  classifier to  the  previous  mode  if  it  was  incorrect, or  maybe  reset  it  to  walking,  actually. And  so  between  those  two  things, the  researchers  never  needed  to  intervene. In  the  experimental  protocol. So  it  was  always  either  the  user's  user  self  correcting through  the  backup  logic  or  through a  manual  cue  through  hip  abduction. So  we  also  had  two  different  types  of  experiments. So  one  was  self  paced, so,  you  know,  comfortable  walking  speeds, just  going  back  and  forth  between  seated  position, walking  in  ramps,  sorry,  in  stairs  and  back. And  so  this  is a  more  typical  protocol where  we  just  get  a  lot  of  back  and  forth  trials. And  also,  you  see  here, we  do  both  intact  leading and  prosthetic  leading  transitions, both  for  ASN  and  decent. Okay.  Alright.  And  you  can  see  the  TV  in the  background  is  just showing  what  the  leg  is  thinking  is  happening. And  then  so  that's  the  self  paced  results. And  then  we  also  have  what  we call  rapid  pace  endurance  trials, which  is,  I  think,  more  of  an  uncommon  protocol, but  really  important  for  showing  that  this  could  work  in the  real  world  because when  people  are  well  rested  and  focused, you  know,  the  results  may  differ  than  if  they're fatigued  and  exhausted  after  walking  all  day  long,  right? And  in  fact,  some  prior  studies  have  shown that  classification  errors  can  go  up  with  time. And  so  we  did a  really  gruing  set  of tests  where  they  did  this  activity  circuit, which  includes  the  ramps. And  we  had  essentially  they  had  to  keep  up  with a  certain  pace  that  was timed  based  on  their  maximum  safe  walking  speed. And  if  they  couldn't  keep  up  with that  pace  for  three  laps  in  a  row, they  consider  that  a  failure, and  then  the  experiment  ended. So  essentially,  as  they  tired,  you  know, they'd  be  more  likely  to  either  make errors  just  slow  down  because  they're  tired. And  so  we  see  one  participant  actually made  it  to  the  very  end  of  our  two  hour  experiment. We  had  to  cut  them  off. That  was  3,000  3,000  walking  steps  in  particular. And  you  see  you  have  almost  500  Sita  stands  and  so  on. So  so  one  of  them  was  very  impressive, and  the  other  one  did  eventually fatigue  and  did  end  the  experiment, you  know,  before  the  cutoff  time. But  what's  interesting  is  that  we  did  not  find any  decrease  in  classification  accuracy. In  fact,  if  anything, classification  accuracy  got  better over  time  because  they  got  in  the  zone, and  they  were  able  to  make  more consistent  motions  and  so  on. So  that  was  interesting. And  then  I  already  talked  about  the  recovery  rate  that they  were  able  to  they  did  have  misclassifications, they  were  able  to  fix  them  on  their  own. I  will  point  out  the  caveat  though  that they  did  have  handrails  and  stuff. So  maybe  it's  easier  to  deal  with those  misclassifications  when  you can  lean  on  something  and, you  know,  maybe  they're  not  quite to  the  point  where  they  could  just  go out  the  real  world and  deal  with  those  misclassifications  yet. So  we've  been  trying  to  expand our  impact  by  deploying  this  on  the  open  source  leg, which  is  a  project  spearheaded  by  Elliott  Rouse, my  colleague  at  Michigan. And  so  we  have  provided  some  ready to  use  controllers  on  opensource leg.com  that  you  can  load onto  your  open  source  leg  right  now  and  use. And  so  that  is  the  sit  to  stand  and  walk  functionality. And  more  recently,  we've  been  trying  to do  a  transfer  of  these  control  methods  to  the  OSR  power. So  we  were  collaborating  with  OSR. They  were  kind  enough  to  send  us the  hardware  and  give  us  access  to  the  firmware so  we  could  reprogram  the  controller  using  our  methods. And  so  we  just  completed  this  study with  seven  above  knee  amputee  participants  to  investigate the  clinical  benefits  of  our  controller  compared  to the  stock  controller  and the  users  take  home  passive  prosthesis. So  this  shows  a  demonstration  that the  same  basic  functionality  exists for  little  choppy.  I'm  not  sure  what's  going  on  there. Between  the  two  devices. So  this  is  on  the  left  is  our  in  house  hardware. Much  bulkier,  less  pretty compared  to  the  Power  k  on  the  right. Now,  what's  really,  I  think, unique  about  our  approach is  that  it  has  this  phase  variable  encoded  into  it. And  I'm  going  to  down  a  bit. And  so  what  you're  seeing here  is  that  as  he  opens  the  door, he  needs  to  move  his  leg  back  out  of the  way  to  avoid  hitting  the  door,  right? And  so  that's  his  hip  motion, driving  the  phase  variable, which  then  tells  the  swing  phase  to  back  up. So  actually,  this  is  a  backward  progression through  the  gate  cycle  that  allows  him  to open  the  store  pretty  conveniently. And  then  we  can  also  do  things  that were  not  explicitly  designed, such  as,  you  know, crouching,  tying  your  shoe,  and  then  standing  up. This  is  all  just  the  sit  to  stand  controller. Alright,  so  in  terms  of  results, this  is  a  quick  preview of  some  of  our  findings  with  this  study, which  you're  the  first  to  see  it. So  this  is  we  increased so  the  HKICPower  again  is our  controller  on  the  OSR  Power  knee, which,  again,  the  Power  Nne  is a  commercially  available  device. Um,  so  we  increased  the  number of  repetitions  during  a  32nd  sit  to  stand  test. So  the  idea  is  that,  you  know, do  as  many  reps  as  you  can  sit  to  stand  30  seconds. And  then  we  repeated  that  ten  times  ten  sets. And  so  we  increased  the  number  of  reps  by 12%  compared  to  the  passive  leg, and  increased  it  by  29% compared  to  the  stock  power  new  controller, which  does  mean  that  the  stock  controller was  slower  than  the  passive  condition. We  also  reduced  ground  reaction  force  asymmetry between  the  sound  side  and the  prosthetic  side  by  31%  compared  to passive  and  14  5%  compared  to  the  stock  controller. And  what's  interesting  is that  if  you  look  over  time  here, this  is  the  X  axis  and  set  number, so  time  by  sets,  discrete  time,  I  guess. We  see  that  actually, they  improved  their  performance over  time  with  our  controller. Which  comp  whereas  the  other  conditions they  either  stayed  flat or  actually  got  a  little  bit  worse. I  don't  know  if  that's  just  learning  and  becoming more  familiar  with  a  new  controller over  time  or  if  it  has  something  to  do  with  fatigue. I  haven't  really  teased  that  out, but  it  was  an  interesting  observation  that was  unique  to  the  HKSC  controller. For  walking,  we  showed  that  for,  you  know, clearing  for  a  fixed  speed, we  were  able  to  increase  toe  clearance  by  23% compared  to  passive  and  increase it  by  6%  compared  to  the  stock  controller. And  then  we  also, I  think  this  is  one  of  my  favorite  results, we  reduced  the  peak  hip  flexion  moment by  large  amount  by  27%, compared  to  passive,  23%  compared  to stock  Power  knee  and  that's  what  we  see  in  the  plot  here. And  why  that  matters  is  that's  that  swing pull  off  moment  to  initiate  swing. And  typically  an  amputee  gait  above  knee  amputee  gait the  patients  have  to  use excessive  hip  flexion  to  cause the  knee  to  bed  B  if  the  pass  the  device, the  knees  not  going  to  bend  itself,  right? So  they  have  to  really  kind  of  whip their  prosthesis  to  cause  the  knee  to  bend. And  that  is  an  overuse  of  their  amputated  side  hip, right,  which  can  lead  to  hip  pain,  back  pain. And  so  we  were  able  to  reduce that  moment  by  a  decent  amount through  what  we  think  is  through better  energy  injection  and synchronization  with  the  gait  cycle. Okay,  so  that's  the  preview  of  the  PowernRsults. And  so  now  just  to  change  topics  to  exoskeletons, we've  also  been  working  on  partial  assist  exoskeletons for  broad  populations. So  we're  working  on  modular  devices that  can  be  used  at  different  joints or  different  patient  populations, but  focusing  on  individuals  with  mild  to moderate  impairments  or actually  maybe  no  impairment  at  all, because  it  could  be  just  fatigue  due to  repetitive  tasks  like  in  a  warehouse. And  you  can  see  that the  numbers  here  are  huge  in  terms  of the  implications  for  partial  assistive  exoskeletons. And  so  our  objective with  the  hardware  was  that  it  should  be  lightweight, should  be  back  drivable, which  means  that  the  human  joints can  cause  the  robotic  joint  to  move  freely,  right? So  we  want  to  facilitate  and support  volitional  motion  because we're  dealing  with  mild  to  moderate  impairments. And  so  vectorability  is  really  key  for  that. And  we  also  want  Comfort  is  an  obvious  one, and  then  generalizable  or  modular, meaning  you  can  apply  it  to  different  joints in  different  use  cases. And  so  we  have  what's  called  the  M  Blue  system, which  is,  you  know, a  fun  name  to  have  M  and  blue  in  it,  right? And  this  is  based  on  an  off  the  shelf  actuator, which  has  now  become  quite  common. I  know  that  Ans  group uses  it,  and  other  groups  do  as  well. The  T  motor  AK  80-9, it's  a  fantastic  motor, has  an  integrated  planetary  gear set  with  a  nine  to  one  gear  ratio. That  can  provide  15  to  30%  of the  biological  torque  of the  joint  depending  on  what  joint  it  is, how  heavy  the  person  is,  what  activity  they're  doing. So  it's  a  range.  And  then  we  have these  different  modules  where  the  hip  and  the  knee are  built  from essentially  our  aftermarket modifications  of  commercial  braces. So  we  buy  a  T  Scope  brace  from  Amazon, and  then  we  replace  the  joint  and  the  uprights  with some  sheet  metal  parts  so  we  can either  laser  cut  or  water  jet. And  so  the  assembly  process  is  very  fast  and cheap  and  so  for  the  controller, our  objective  was  to  have  it  be  versatile, a  able  to  adapt  to different  activities  to  be  predict  predictable  and safe  so  that  it  doesn't  have unexpected  oscillations  or  unexpected  behaviors. We  want  it  to  be  generalizable, customizable  to  different  users, and  we  also  of  course,  should  be  effective,  right? Should  be  effect  at  something,  whether  that's  reducing effort  or  improving  performance  or  mitigating  deficit. And  so  our  approach  to  doing  this  is  we  take  a A  non  linear  controls  or  dynamical  systems  approach, which  is  related  to  passivity  based  control. So  the  idea  is  that  we  aren't  going  to  enforce any  trajectories  because  trajectories would  inhibit  volitional  motion. We  don't  want  to  prescribe  the  joint  pattern  of  the  user. Their  own  joint  should  be  doing  that. Instead,  we  want  to  change the  dynamics  of  the  body to  make  it  easier  for  them  to  move. Reduced  effort,  right? And  so  if  you  think  about  the  human  robot  system, the  coupled  system  here,  you  can  model it  through  Lagrangian  function, which  through  the  oil  and  Lagrange  equations, gives  us  these  canonical  equations, which  I'm  sure  you  can't  see  back  there, but  you  can  check  out  papers  if  you'd  like. Um,  and  then  we  do  is  we  have access  to  this  control  input at  these  joints,  for  example. And  so  we  close  the  feedback  loop  with a  control  law  that changes  the  closed  loop  dynamics,  okay? So  once  you  close  the  loop, the  dynamics  are  now  altered. So,  for  example,  maybe  we've reduced  the  perceived  gravity  of the  human  or  perceived  inertia or  stiffness  or  something  like  that. And  so  this  can  be  done in  an  energetically  passive  manner, which  means  that  in  the  continuous  time  dynamics, you're  not  injecting  energy, but  you  can  inject  energy through  the  hybrid  dynamics  through  switching. For  example,  if  you  switch if  he's  changed  the  set  point  of  a  virtual  spring,  right? You  can  inject  energy  that  way, but  it  would  be  like  a  transition  of, like,  going  from  stance  to  swing. Okay.  And  so  it's  locally  passive, which  has  benefits  in  terms  of  safety  and  predictability. And  so,  examples  of  the  types  of  features  that  we  can design  into  this  framework  are  gravity  compensation, inertial  compensation,  and  virtual  springs and  dampers  as  well. So  now,  we  have  a  relaxed  version  of  passivity, as  well  that  allows some  energy  injection  in  the  continuous  time  dynamics. I'm  going  to  kind  of  skip  over  that, but  we  do  have relaxed  forms  of  this  that  aren't  so  strictly  passive, and  that  has  helped  us  do  things like  predicting  human  joint  torques  from  biological  data. So  what  we  do  is  we optimize  over  an  energy  shaping  basis, so  essentially  a  set  of  nonlinear  functions  to  predict normative  joint  torques  given normative  kinematic  or  ground  reaction  force  inputs. And  this  is  over  a  multi  activity  dataset. So  you  can  see  here  that  we  have  ramps,  we  have  stairs, we  have  level  walking  and  sit stand  and  for  the  knee  and  for  the  ankle. And  so  essentially,  what's happening  is  instantaneously  we're mapping  the  kiomatics  and  ground  reaction  forces to  the  biological  joint  torque or  at  least  our  best  estimate  of  that. And  so  that  way,  then  we  can  provide  a  fraction  of that  biological  joint  torque  to  the  human to  reduce  their  effort  at  the  joint. And  so  this  is  End  to  end. It's  not  in  the  same  way  as  a  neural  network  is, but  it's  end  to  end  in  terms  of  taking  inputs, directly  mapping  it  to  outputs. There's  no  classification  involved. So  and  so  you  can  see that  it's  a  modular  framework  as  well. So  we  have  a  modular  basis that  can  be  used  on  different  joints. I  know  there's  a  lot  of  videos  going  on  here, but  you  can  see  it's  a  hip, we  have  a  knee,  we  have  an  ankle, we  have  unilateral,  we  have  bilateral. And  so  this  framework  can  be  used  to design  any  joint  configuration. And  so  then  we  had  a  recent  study  using  this  framework  or at  least  a  control  method  that's  inspired  by this  framework  to  try  to mitigate  fatigue  during multi  terrain  lifting,  lowering,  and  carrying. And  so  the  objectives  here  were  using  a  knee  exoskeleton. We  wanted  to  reduce  the  contributions  to quadriceps  fatigue  over  this  multi  terrain  lifting, lowering,  carrying  circuit. The  idea  there  that  if  we  can  reduce the  contribute  if  we  can reduce  the  contributions  to  quadrocep  fatigue, maybe  we  can  prevent  or delay  fatigue  in  the  first  place,  right? But  why  do  we  want  to  do  that? Well,  because  when  people  are  fatigued, they  tend  to  take  riskier  lifting  um  lifting  postures, like  scooping,  lifting  with their  back  instead  of  their  knees,  right? And  then  people  get  injured. So  we  want  to  prevent  fatigue in  the  first  place,  if  possible. But  then  let's  assume  that  you  are  fatigued. Let's  say  you've  already  gotten  there  because  you've been  working  in  the  warehouse  for  your  eight  hour  shift. Then  are  we  able  to  mitigate the  deleterious  effects  of  fatigue? Okay?  So  we  tested  both  of  these  things  in  this  study. And  so  first  to  the  first  objective, we  focused  on  quadriceps  EMG  effort, which  is  essentially  the  integral  of muscle  activity  averaged  across all  the  quadriceps  muscles  that we  measured,  three  muscles. And  so  we  were  able  to  show that  um  we  were  able  to  show  that  for  all  the  activities, the  average  muscle  effort decreased  with  the  exoskeleton compared  to  no  exoskeleton, but  that  was  not  statistically significant  for  level  walking. It  was  statistically  significant for  all  the  other  activities. It's  not  surprising  that  it  wasn't significant  for  level  walking  because the  knee  the  quadroceps  don't do  very  much  during  level  walking  anyway. So  the  baseline  is  so  much lower.  It's  hard  to  reduce  that. So  we  weren't  surprised  by  that, but  the  real  benefit  is  during  the  harder  activities, like  lifting  lowering  and  assent decent  tasks  where  reducing  hydrogen  effort  is  valuable. And  then  this  video  shows a  repetitive  lifting  and  lowering  fatiguing  task. So  the  idea  here  is  that  we  had  them. It  was  not  a  fun  experiment. We  had  the  participants  do  repetitive  lifting  and lowering  as  quickly  as  they could  until  they  couldn't  do  it  anymore  until  failure. So,  you  know,  their  quads  are  burning. They  can't  do  it  any  further. And  then  and  then  we  have  on  the  left, we  have  not  wearing  an  exoskeleton,  and  on  the  right, exoskeleton  is  donned,  but  it's  passive. It's  in  its  zero  torque  mode. So  it's  not  helping  maybe  resisting  ever  so  slightly. We  also  accounted  for  that  in  the  analysis. And  then  once  they  declare failure  that  they  can't  go  any  further, then  we  give  them  a  second  to  compose  themselves, and  then  we  activate  the exoskeleton  and  we  say,  Okay,  go  again. Now  do  ten  reps.  Do  ten  more  reps  as  fast  as  you  can. And  you'll  see  in  this  video  that  the  exoskeleton, once  we  turned  it  on,  did  really  help  them. So  this  is  the  fatiguing  process. This  is  the  not  fun  part. Takes  a  while  for  some  participants. So  here  they're  beginning  to  really  hurt. And  then  they  say,  Okay,  no  more. Okay,  so  they  can  pause  and compose  themselves  until  they're  able to  do  the  next  rep  with  proper  form. With  exocale  it's  inactive. The  participants  a  little  more  confident, so  they  maybe  don't  pause  as  long. And  then  they're  able  to  knock  out those  ten  reps  very  quickly. All  right,  so  just  for  the  sake  of  time, this  is  a  easier  way to  look  at  that  in  terms  of  the  data. So  this  is  the  complon  time  increase percentage  on  the  left  plot. And  you  can  see  that,  you  know,  some  participants  had, you  know,  this  is a  percent  increase  relative  to  baseline, relative  to  pre  fatigue. Okay.  And  you  can  see, some  participants  understandably  had a  very  large  increase  in  completion  time to  do  the  lifting  lo  cycle  when  they  were  fatigue, right?  Very  understandably  so. With  the  exoskeleton  on, they  actually  return  to their  pre  fatigue  performance  level  or  very, very  close  to  it. I  think  it  was  off  by,  like,  one  or  2%. And  so  at  least  in  performance, we  were  able  to  mitigate  the  effect  of  fatigue. But  also,  we  looked  at  the  postural  changes  as  well. And  so  the  amount  of  stooping, in  particular,  is  of  interest. And  we  were  able  to  show  that  they  were  able to  reduce  the  amount  of stooping  with  the  exoskeleton compared  to  no  exoskeleton  after  fatigue. So  that,  again,  guarantee that  they  wouldn't  injure  themselves, but  it's  one  of  those  things  that  suggests  that the  lifting  posture  is  a  safer  more  proper  form. Okay,  so  now,  wrapping  up, I  want  to  talk  about  some  of our  ongoing  work  that's  hot  off  the  press. So  we  just  got  a  paper  accepted  at ICR  related  to  reducing  pain  and  kneeostearthritis. And  so  hopefully  I'll  see  all  of  you  at  ICR  in  Chicago. And  so  we  did a  four  participant  pilot  study  where these  individuals  had  chronic  kneeoseoarthritis, but  a  specific  form  of it  called  patepemeral  kneeosteoarthritis. They  may  have  so  the  knee  has  three  compartments, the  patepemoral  compartment  is essentially  the  kneecap  area, and  existing  OA  braces don't  really  address  pain in  that  compartment  of  the  joint. That  compartment  pain  is  exacerbated  by  quadriceps  force because  it's  sutily  pulling  a  tendon across  spanning  that  kneecap. And  so  so  traditional  OA  braces  can  only do  mediolateral  adjustments  to deal  with  the  mutilateral  compartments, but  they  don't  help  with  patel  femoral  pain. And  so  here  we  targeted patients  who  have  diagnosed  patel  femoral  osteoarthritis. They  may  have  additional  compartments  inflamed  as  well, but  they  at  least  had  patel  femoral and  it's  an  unaddressed  population. And  so  you  can  see  here  that  over  all  these  activities, the  primary  activities  of  daily  life, we  were  able  to  consistently reduce  pain  reported  is  perceived  pain  of the  user  with  the  exoskeleton  compared  to  no  exoskeleton. And  the  hypothesis  here  is  that  we're  reducing that  quadriceps  force  that causes  a  load  on  the  Patel  premal  compartment. And  we're  also  able  to  show  through  inverse  dynamics  that we  reduce  the  joint  moment  as  well  at  the  knee. So  that  helps  explain why  the  pain  reduction  is  happening. And  then  we  also  extended  this  to the  hip  with  a  pilot  study  of  three  participants, where  a  very  similar  effect  was  observed  that  pain  was consistently  reduced  to  cross activities  and  with  exoskeleton. And  we  have  a  similar  hypothesis. I  mean,  the  hip  is, of  course,  different  than  the  knee. The  sources  of  pain  are  a  little  different, the  hip  is  a  ball  and  soca  joint, so  it's  very  complex. But  here, we're  just  providing  flexion  extension  assistance. So  we're  targeting  just  one  degree  freedom  of  the  hip. But  that  alone  was  enough  to  reduce pain  considerably  for  our  participants. So  excited  about  sharing  those  results  at  ICR. And  then  we  also  have  my  last  slide. We  also  have  an  ongoing  clinical  trial on  mitigating  age  related  weakness. And  so  this  is  also  known  as  sarcopenia  frailty, where  this  individual  in  particular  is  unable to  do  a  step  over  step  stericent  gait  on  the  stairs, which  is  the  more  natural  gait because  of  quadricepts  weakness. So  he  has  to  do  a  step  to  step  gait. So  you'll  see  this  kind  of, you  know,  shuffling  up  the  stairs. And  this  is  his  first  try  with  the  Nooskeleton. And  suddenly  he's  able  to  do  the  step  over  step. So  that  was  a  very  cool  quote  from  him. He  was  excited  about  it  because he  hasn't  been  able  to  do  that  for  many  years. So  that's  an  ongoing  study. Hopefully  we'll  have  some  papers  on  that  soon. And  so,  yeah,  in  conclusi, I  just  want  to  thank  you  for  your  time. And  if  you  have  any,  you  know, interest  in  seeing  this  work  in  more  detail, we  have  a  YouTube  channel  as  well  as  the  lab  website. I  just  want  to  thank,  of  course, the  students  and  postdocs  who  made  this possible  and  happy  to  answer  any  questions.  A. Okay,  I'm  gonna  be  running  around  with this  little  microphone.  This  is  what  we  have  today. So  if  anybody  has  any  questions,  raise  your  hand. I  see  Veron,  then  I'll  come  over  here. Oh,  thank  you  for  the  talk,  by  the  way. When  you  collect  data  from the  human  participants  and  use  that  to train  the  trajectories  of  the  prosthetic, how  do  you  normalize  for,  like, the  able  bodied  participants  having different  body  geometries?  Yeah. Okay,  so  with  impedance, we  can  normalize  the  impedance  parameters  by  body  mass. Torque  is  sensitive  to,  like,  body  mass,  for  example. But  the  kinematics,  like  the  joint  angle  trajectories, those  are  not  sensitive  to  body  mass. I  mean,  of  course, we  all  have  different  kinematic  patterns, but  that  you  cannot  regress  as  a  function  of  body  mass  or height  or  other  geometric  properties  that  I  can  measure. So  you  just  stick  with  the  average  joint  trajectories, and  it's  a  good  starting  point. It's  not  optimal  for  anyone, right?  But  it's  a  good  starting  point. We  do  have  a  clinical  tuning  interface that  lets  us  using,  like, um  um,  we  can,  like, drag  points  on the  trajectory  change  things  do  it  that  way. You  could  also  just  tell  them  you  can tell  the  clinical  interface  also allows  clinicians  to  increase stiffness  at  Heelstrike  and  features  like  that. So  we  have  a  way  of  tuning  it,  but  at  the  moment, we  don't  have  a  good  way  of  making  it  subject specific  just  through  their  geometric  properties. Okay,  you  quit.  Thanks  for  sharing  your  research  today. I  just  wanted  to  ask  about  the  um  uh, what  do  you  call  it  the  lifting test  you  did  with  the  exoskeleton. Would  you  see  any  benefit  in  using like  a  placebo  group  where  you tell  them  that  the  controller  is  active, but  it's  not  to,  like, separate,  you  know,  the  physical  and  mental  aspect. Yeah,  yeah,  improvement  in  exercise. Yeah,  it's  a  great  question, and  I  will  freely  admit,  like, this  is  not  a  perfect  experiment. It  is  really  hard  to,  like, control  for  all  these  things because  people  know  when  they're  wearing  an  exoskeleton, to  answer  your  question  directly, they  can  tell  when  the  exoskeletons  doing  something. There's  no  hiding  when  there's  a  torque  applied  to  you. So  we  thought  about  that. Yeah.  And  you  might  also  wonder,  why  didn't  we,  like, have  them  fatigue  without the  exoskeleton  and  then  put  the  exoskeleton on  so  that  the  fatiguing  process  was the  same  two  sides  between  the  two  conditions. The  reason  for  that  is  that  it  takes like  2  minutes  or  3  minutes  to  don  the  thing. And  by  then,  the  acute  fatigue  has  passed. And  then  they  can,  of  course, they  may  have  longer  term  fatigue,  right? But  it  was  a  really  hard  experiment  to  design,  yeah. And  yeah,  there's  potential  sources  of bias  due  to  their  knowledge of  what's  happening.  That's  true. Good  question.  Another  one. Professor,  I  just  had  a  question  about you  talked  about  lifting  posture deteriorates  as  people  fatigue  when  lifting. Did  you  mandate  a  certain lifting  posture  when  they  started? Yeah,  yeah,  we  said  they  had  we  talked  to  them  about the  difference  between  a  squat  lift  and  a  stoop  lift. We  told  them,  maintain  a  proper  squat  lift  with  your, you  know,  back  as  vertical  as  possible. And  so,  yeah,  we  did  have  a  Oh, one  thing  I  forgot  to  mention  is  that every  participant  in  this  study  had prior  experience  with squat  lifting  in  the  gym,  for  example. They  weren't  necessarily  warehouse  workers, but  they  were  people  that  had  done  squats  in  the  gym. They  were  familiar.  That  was  our  best  way of  getting  closer  to the  population  that  matters  for  this  particular  use  case. Go  cha,  can  I  ask  a  quick  follow  up  question? So  when  you  assist  at  the  knee, do  you  think  if  you  didn't  cue for  a  person  to  have  different  lifting  posture, do  you  think  that  assistance  profile  will  change  how a  person  default  wants  to lift  using  their  knee  versus  using  their  back? So  you're  talking  about  the  system profile  from  the  exoskeleton. Yeah,  let's  say  that  we're  an  exoskeleton. Do  you  think  that'll  change their  lifting  form  because  they  have a  different  kind  of  energy expenditure  at  the  different  joints? Well,  it  did  change  their  lifting  form  in  fatigue. But  if  I  go  back  here,  let's  see. Did  it  change  them  before  fatigue? Uh,  kind  of  hard  to  say,  right? So,  this  negative  region of  the  trial  is  pre  fatigue or  before  they  declared  fatigue. Zero  is  when  they  declared  fatigue. I  probably  should  have  explained  that. Anyway,  this  is  before  they  declare  fatigue. And  so  there's  not  a  lot  of  difference  in  the  amount  of stooping  between  the  two  conditions. But  then  again,  it's  not  off. So  it's  off.  Okay,  yeah,  yeah,  yeah,  yeah. I  guess  I  don't  have an  answer  to  your  question  with  Exelon. Yeah. Thank  you.  I  saw  this  one  next, and  then  I  Thank  you.  All  right. My  question  was  about  the  ICR  paper  with  the  knee xo  might  be  a  bit  quicker, but  were  there  any  modifications  to the  controller  between  lift  load carry  and  reducing  pain  in  this  case? Yeah,  so  we  did  modify  the  lifting lowering  functionality  to  be better  for  sitting  and  standing? They  actually  are  very  similar. The  range  of  motion  of  the  needle  is  similar  for  lifting lowering  versus  a  squat  lift  versus  a  sit  stand. So  we  did  make  some  modifications  for  that. And  there  might  have  been one  other  modification  is  eluding  me. I'd  say  these  are  minor  modifications. I  would  not  call  that  the  contribution  of  the  paper. But  what  was  nice  is  that  the  controller  was customizable  to  this  population, which  I  think  worked  to  our  benefit.  No. And  you  one  more  question.  I'm  sorry  if  I  missed  anyone. This  might  be  coming  from unfamiliarity  with  ne  osteoarthritis, but  is  it  usually  bilateral  in  the  sense  that  were  you giving  assistance  to  the  same  extent and  did  they  have  pain  to  the  same  extent? Do  you  know  if  that  would  affect  things? Right,  so  yeah, some  patients  have  bilateral  OA  and  some  don't. And  so  what  we  did  is  we  put the  exoskeleton  because  it  can work  actually  we  use  bilateral. We  use  a  bilateral  exoskeleton  for  all  patients  because the  controller  at  the  moment,  involves  bilateral  sensing. I  do  have  a  student  working on  a  unilateral  version  of  that. But  for  this  study, it  was  bilateral  for  all  patients, but  some  only  have unilateral  OA  pain.  Does  that  answer  your  question? Yes. Oh,  okay.  Sure. Yeah,  yeah.  So  you're  saying, maybe  an  exoskeleton  can  help  compensate  for the  overuse  of  sounds  or  intact  joints,  right? It's  a  really  interesting  point,  yeah. But  yeah,  if  you  use  an  exoskeleton  for  that, you  could  in  theory, have  the  prosthesis  communicate  with  exoskeleton, and  then  you  can  do  some  really  cool  stuff then.  That's  interesting. Thank  you,  everyone.  Let's  thank  the  speaker. If  you  have  more  questions,  maybe  just  ask  him  after, but  thank  you.  Yeah. Um.