Well get started.

It's very easy for me to introduce
the speaker today as I'm the speaker so

I wanted to talk a little bit
to you about not necessarily so

some mixture of what some of the research
that's going on in my group but

also I want to talk
a little bit to you about.

Some of the take away lessons for
what's happening in manufacturing and

in in automation and try and argue that.

Some of you might have the view that
automation and manufacturing is really

boring and there's actually some really
cool stuff going on in manufacturing and

so I want to talk a little bit about
that so I want to talk I want to start

out by talk a little bit about what
are some of the future trends and and

talk about why how is that sort of shaping
our industry in terms of doing robotics so

the good news is that there's a lot of.

Different kinds of reports out there
that are trying to point to why is.

Why is manufacturing
becoming a big deal and

how are we seeing sort of this
integration of I.T.N. manufacturing into

a new area it was a very nice McKinsey
report that's really talking about how

the global economy is shaping what we do
in terms of future manufacturing is a very

nice also report about how is why is
manufacturing coming back to the U.S. So

if you look at it really a noble reason to
build a factory in North Carolina Motorola

build a factory in Texas so we've seen all
of these companies coming back to the U.S.

and it's a very nice report also about.

From the Atlanta Council
call alternate worlds or.

And really talking about what would happen
if trying to outline what might happen by

twenty thirty white so
what would happen if.

China crashed what would
happen if China had

economic dominance over the U.S.
So it's trying to.

Line a number of different scenarios and
then it's not saying one of the other

is to write one but it's trying to sort
of outline saying which one of these why

you want to think about so it's a great
way of doing this of course the good news

about is that is that it's easy to
write these reports it's much harder

to get the truth about how can
vacuum predict the future so

one of my fellow countrymen I think it was
in the foresight you know the hardest part

about making it's hard to make predictions
especially about the future and

some people have done sort of blatant
errors in this so Thomas Watson when

the first computer came up so to say that
the world market for computers is for.

He was off by some orders
of magnitude when.

The first computer came out
the CIO deck came out and

say there was no market for
these computers in the home.

He was also a little bit off he only
thought about that they could be used for

spreadsheets and who would want
to spread sheet in the home and

when the telephone came out Union
wire basically the Union Telegraph

send out a memo in saying using the phone
is not a convenient modality for

human communication so
they will all fall by a little bit so

sometimes it's very hard to sort
of see where is the future and

what are some of the some of
the interesting opportunities out there.

I want to talk a bit about what are some
of the mega trends I'll talk about where

are some of the the things that
sort of really changing the welt

put a little bit about in R.

and D. and I'll talk about some of
the example research that's going on.

In my research group so so first talk
about the mega trends one of the big

mega trends that's driving a lot of
this is extreme cost to my sation so

we all want products that are our
products they are unique to me.

So why did I put up a jaguar a Jaguar is
one of the simplest cars you can buy.

From a configuration point of view

it's only available in one point six
million different configurations.

So we all want different
audience different color

different wood panels
that from leather G.M.

in theory is manufacturing ten to
the twenty third different com audibles.

That's a really serious challenge
to be able to do automation and

that's why if we go and look at sort of
the automotive industry we've basically

automated the plate shop so
you bring in a piece of them in there and

you punch in sort of into what
this ass is supposed to look like

we've done the welding shop and
we've done the paying shelf or

ninety percent automated once you go
beyond this you stop the customization and

because you stop the customization this
very little automation taking place

we're starting to see a few
places where it's getting out but

it's been very hard simply because we love
to get these we love to get for the mass

customization it's going to continue to do
that and that really sort of challenges

some of the things that we do the other
thing is that we're starting to see.

Mass customization of electronics
would sort of challenges us so

if you think about it I don't
have my latest i Phone But

you know the latest i Phone just came out
Apple is already working on the i Phone

seven already deciding the automation
system for the i Phone seven it will

come out in ten months so the problem is
that how do you build an automation system

that has a life time that might
be if you're lucky a year and

you know the good news is that the robots
you saw here typically has a life time of

ten years but the product they're making
might only have a life time of months and

that's sort of really a challenge is how
we can design new automation solutions

they have to be efficient enough that
you can spend years programming them so

they need to be able to be
programming almost one else so

we're seeing this we're seeing a lot of
automation coming back to the U.S. but

it does require us to be able.

To do much better programming than we
do today you can afford to spend a lot

of time on doing this and
of course we're starting also to see

Smart Cars are starting to see a Thomas
driving car some of you might see that

there's a lot of press right now in in the
state of Georgia about how can we be able

to autonomous driving cars how can we
allow them to drive on the road so

there are four.

States in the U.S. right now where
you're allowed to drive them so

it's California and
Nevada Florida and Michigan.

And and there we're seeing and there's no
doubt we will see these my own prediction

is that children born today
will never have to drive a car.

The prediction from Google is that
will happen before twenty twenty

I think they're being overly optimistic so
I think it's going not because I

think we can get the technology wise
we can get the lawyers to do this so

the lawyers are going to break us once
again simply because it's very dangerous

on the other hand I feel very safe in
these not too long ago I drove across

the Golden Gate Bridge
the Google car on one side and

I was you know the guy in there was
reading linear algebra and I was like OK.

That was his choice on the other
side I had a a young woman

talking on the phone drinking coffee
trying to apply makeup at the same time

I was much more worried about the young
woman than I was about the Google car and

so you know we we're all very
distracted drivers we have so

much technology in our car that easy for
us to get distracted so getting

an autonomous driving car will be a big
deal and we'll see this all come in and

talk about how this sort of
impact some of our logistics.

One of them and one of the reasons
why we're seeing this is of course

e-commerce is right now growing forty
percent a year manufacturing is growing

eight to ten percent it's honorable but
it's not going forty percent so one of

the reasons why Google is very interested
in building autonomous driving cars is.

Is to be able to do automated custom
delivery already today you can

see from Amazon in big metropolitan areas
and we are apparently not one of them

you can get Amazon Fresh You can order
your groceries on the way up and

you can have them delivered to your home
by five P.M. By the time to get home

in my neighborhood it would be
gone by the time I get home but

you know that's a different question and
all but but Amazon can deliver to the you

the big challenges that Google would
like to be able to compete with

Amazon in the space not because they're
really interested in robotics but

because they are worried about about five
percent of their advertising revenue

disappears to Amazon so Amazon you know if
you're going to go and buy a new pair of

sneakers or whatever you will go straight
to Amazon you will no longer pass through

Google because of this they lose five
percent of their advertising revenue five

percent of advertising revenue for Google
is a truckload of money for that reason

a very interesting getting into the space
you're seeing the same competition in

a share between Alibaba and buy do

also very interesting trying to understand
how can they really get into this space.

Here is you know one of the challenge is
this looks like a really nice way to house

the challenge is that.

Thirty percent of the space
is filled up with stuff and

seventy percent of the space is filled up
with the air if you live in Atlanta it's

not a big deal if you live in D.C. Boston
and San Francisco that's a big deal space

is expensive so we're starting to see
robots being deployed to do this you might

have seen Kiva come out building these
a lot scale systems they were sold for

seven hundred million there's a company
we worked with symbolics out of LOS them

that have built warehouses where they have
ten stories high robots they have four

hundred robots deployed in this system
to be able to get twice the packing

efficiency twice the packing
efficiency is a lot of money in this so

you will see lots of robots being
deployed in these warehouses for doing.

What was called automated stories and
retrieval systems so

what we're seeing is that Emma Fed Ex and
U.P.S. and those are very interested in

how can they get automated cars to be able
to do the livery they've done that they've

worked out the business case they still
have a guy in the back of the truck but

they don't need a guy in front of the
truck and it makes business sense because

you don't have to pay the guy
to have a driver's license.

So that scary you know soon you'll
be a monkey sitting in the back

running around with baggage is
why somebody else drives the car.

We saw you know and this for for the White
House logistics at the height of the war

in Iraq and Afghanistan the U.S.
had fifteen thousand autonomous be flying

vehicles deployed there it was a great
space right now there's a lot of interest

in trying to get automated flying vehicles
for doing automated delivery you've seen

Amazon you've seen Google
work on small quarter orders.

The big market short term is going
to be coast to coast air freight

it's going to be the fact that I can
fly stuff from here to San Francisco

with no pilot you know last week that last
week I was in Hong Kong if you think about

it the pilots spend half an hour sort
of taking off then they spend fourteen

hours playing Angry Birds or whatever they
do and then they land again they face and

you have that pilot sucks and
he has to be away probably ten days

a month from his family and from all sorts
of other things if he could be sitting in

a container somewhere running a fleet of
aircraft so he would have a much higher

efficiency at same time I would be able to
build an aircraft that doesn't need oxygen

it doesn't need to be pressurized to the
same level doesn't have to have a restroom

onboard so I can build much cheaper
aircraft that I could do before so

there's a lot of interest
in being able to do this so

you'll see this very soon think of
take a little bit longer before as

a passenger I would be willing
to step on an aircraft and

you know the pilot says I'm sitting
here on the tarmac in Kansas and

I'm dealing with him and I mean I
glance have what's he doing there.

And say well you know you will notice
there's nobody at the front of the cabin

because front row seats will
be interesting by then.

But it will take a while before
we are ready for this but

to give you a data point forty percent
of the pilots entering the U.S.

military today never get to fly
they become drone operators

forty percent that number really
shocked me that it's that high so

it's a lot of really cool technology
happening out there another area where

we're seeing this isn't time to doing
aerospace manufacturing so both Boeing and

Airbus right now have a backlog
of seven years so when the F.

and decides he needs his own seven thirty
seven to do international business he

exactly so the goodness is you can
probably get three to buy one tomorrow but

you can get it delivered twenty twenty
one so that's a huge challenge that they

just can't deliver and
it's a huge challenge for

them in terms of doing new technology in
terms of doing new new initiatives so for

that reason they are looking can they
reduce their their backlog to about three

years to be able to do this they would
like to double their manufacturing volume

so both Boeing and Airbus who see
an Airbus is building a factory in

Mobile Alabama we're seeing Boeing right
now for the seven hundred seven trying to

increase the volume from thirty two
aircraft a month to seventy two aircraft

a month every aircraft is one hundred
million dollars That implies that they're

trying to increase their turn over
a four billion dollars a month

this is serious money and to do this
they really want to figure out how

can they have much more automation than
they have today how can they automate this

to a much higher degree they're barely
any robots being used in here today so

there's a number of interesting
challenges of how can you do this

another interesting piece of data that
you actually might be aware of but

I was in the way of is that the average
American place first person computer

games about one hundred fifty hours
a year some of you a picking up my slack.

I don't pay for person games at all so
some of you are clearly picking up for

me it's half an hour a day a more
interesting data point is that

the average twenty year old American male.

Has spent twelve thousand
hours playing computer games.

That's free years with without sleep.

You know.

That's an interesting data point because
if I could decide a manufacturing system

that looked like a computer game
these people would be instant experts

I would be useless you and
I would be the yellow button is for

what and who cares you know I'm
not the prototypical user so

we will see people if we could decide
to use innovations that looks like this

that looked like this we would be
able to do some really cool stuff so

it's really nice to sort of think about
how can how can we think about gaming as

being an interface modality because
the people that are entering the workforce

are instant experts and just to give
you another data point that really

depressing data point in
the state of Georgia forty two

percent of a generation of young
people do not graduate high school.

Forty two percent that's of ridiculously
large number the good news is that

Georgia is at the bottom of the Union so
it's Alabama and

Mississippi Georgia Florida
sort of the worst states so

if you go to Boston California all these
other places those numbers are much higher

but if we want for instance to have
a manufacturing plant here in Georgia

this is a significant chance that
they might be a gaming expert but

they don't have a high school diploma
you know so that's a very interesting

sort of data point and how can I design a
manufacturing system that would look like

something that they could all use even
though they can graduate high school

another thing that's really sort of
moving forward right now is salaries so

the good news for for us is that
the United States actually has ready.

If the low salary is you know
Denmark where I come from and

Sweden have much higher salaries
in comparison then of course

there are other countries like Mexico
that's that's even cheaper but

we're sort of in the middle but one of
the things that's been a real game changer

is that salary in China and
in Southeastern Asia

has gone up by three hundred fifty
percent over the last ten years.

Ten years ago you would not hesitate
about moving manufacturing to

China simply because the cost of
labor was so low the fact that it got

up by four hundred fifty percent implies
that it's no longer necessary a good deal

to move your manufacturing to Sofie some
china that's why we're seeing Motorola

all of these coming back and putting their
manufacturing here so there's some really

interesting opportunities to to figure
out how can we compete in this area so

some of the mega trends are really
how do we do mass customization.

How can we do sort of almost
immediate delivery and

doing manufacturing how can we have
how can we manage these very short

product cycles how can we optimize
that sort of all the way.

And another so I want to talk a little bit
about some of the things that's really

changing so one of the things that's
changing is that we've been used to using

these massive machine tools if you want
to see some of them you can go down

we have a number of these for
instance in the in the mock.

You can go over and see these that
characterized by having extreme accuracy

they're very good you can apply very big
forces the challenge is that they're not

very adaptable so if you look at the Lodge
a manufacturing company is if you design

a line using these kinds of equipment it's
very hard to adapted from one generation

to the next I just told you that the
average lifetime for product might don't

to be six to twelve months so
the question is how can I change these so

what we're seeing is that we're seeing
these being thrown out being replaced by

larger robots that will allow us to go in
and do much more efficient processing and

much more data processing so
we can use it over and over again so

we're seeing this for instance for making
very large scale composite machines.

So for a C.

windmill the length of the blade
of one blade might be eighty

meters that's two hundred forty feet long
plaits we're talking about serious here so

when blades they're all being
manufactured by using robots to do this

on some of the lots drilling machines
this is drilling machine for

aerospace the largest trails
are ten stories high.

So you know that poses sort of
a very interesting challenge and

how can I take these monuments and replace
them with robots so we're starting to see

much higher base of automation where you
have here two robots that can manipulate

an object wired into our object by two
other robots are used for doing welding so

having sort of a free to work cloud
of robots you can have a much

higher to be of flexibility and we seeing
this so that now down a conveyor belt

you can maybe have every one of your cars
could be different or every one of your

airplanes could be different and by doing
this you can do some really cool stuff

the big challenge now is how can I
give these robots extreme accuracy and

the accuracy is typically better
than a thousand and seven and.

So and and that poses an interesting
challenge I'll talk a little bit about

this later on of course we're seeing
a lot of work on collaborative robots.

So you've all heard about the back
of the robot the interesting a hard

thing about the max robot is
its performance sucks so so but

the interesting part is it's true it's
only twenty thousand dollars the fact that

we can build a fortune where freedom
robot for twenty thousand dollars and

it becomes sort of a commodity item by the
time it breaks you just throw it away and

you buy a new robot we've seen
interesting robots from universal that

were the interesting part there's really
we've gone to where we have the I.S.O.

standard to ten to eighteen point six that
it designates these robots to be safe and

have a different discussion
about whether they are safe.

But it's very important that we have
the collapse of robots that we can use for

doing this and now they're finally coming
out with robots that have a reasonable

degree of accuracy so the latest
generation of the Universal Robots have

encoders on the output shaft and because
of this we can get down to absolute

accuracy that are in the range of point
two millimeters which is pretty good and

terms of if you think about it and
that's without putting sensing and

of course I would argue we should
put some sense into these systems.

And of course we're seeing this from being
able to do teaching by demonstration I

don't have to talk to you about this and
we're seeing new ways of really thinking

about how can we use the cloud how can
we use the cloud to get very expensive

systems in the U. in the Europe that's
called industry four point zero and

somehow in the U.S. It's called
the industry and Summit same thing but

the idea is really how do you
combine that entire process and

use this to build up interesting
of what you're into this so G.E.

is pushing this where the idea is that
everything from what's local on the floor

all the way up to the boardroom is fully
interconnected on an industrial Internet

and by doing this you can do source up
to my station and do really interesting

things so do you estimates that over a ten
year period we could save sixteen trillion

dollars That's an interesting number.

But it's sort of it allows us to make
very serious investments in doing this

the other thing that we're seeing is that
we're seeing a lot of interest in doing

data analytics so we've seen a lot
of interest in doing finance and

insurance and we have a lot of people here
attack that's very good at doing this but

also we're starting now
to see manufacturing

as an area where we can do significant
data analytics to optimize the process and

really understand how can
we streamline the process.

So there you will see some some
interesting new account is coming out

the other thing we're seeing is that we
are seeing interesting new processes so

that Tesla is a great example of
a car that was designed for to me.

So finally it's a sexy sports car
you know I drive a Porsche so

when they told me you can have
the military Yoda I was a like really guys

now I'm thinking about this is a sexy car
but the real thing that's interesting

about it is it's built in California which
is the most expensive place in the U.S.

to do manufacturing and they did this
by deciding the car day one to be

used with a very high degree of automation
so you can use this very high degree of

automation so they have something like
forty percent automation compared to or

twice as much as most automotive companies
and they did it partly by thinking

about sort of one hundred years ago
Henry Ford said You can buy your

your car in any color you like
as long as it's black and

Tesla has some have done the same
thing here said we're going to limit

the variability of the number of
cars you can use to a very sort of

fixed said of of different configurations
so they don't have one point six

million configurations they are more like
in the hundreds configuration doing this

they can use robots to a very large
extent to assemble the entire thing and

because of this they can do manufacturing
in California rather than somewhere else.

We're seeing massive use of.

Cheap computing My I would say computing
cost nothing today so this is an example

of an abort that came out about a year
ago from a company called parallel it has

one hundred twenty eight course on
a credit credit card sized computer

it has one thousand meg of one thousand
Giga flops compute power on there and

it costs eighty dolls' So the fact that
we can get compute power from nothing and

do this sort of really make some of the
difference of being able to do this they

promise they will come out with a thousand
and twenty four course this year so

and it's small enough that you know if
you fall away all of the connectors and

all of that we're sort of talking two
dimes in terms of doing compute power so

that opens up a number of
interesting of which are into this.

One of things we've seen if you go and
sort of try to deploy a robot in the real

world is that typically twenty percent of
the cost of the robot twenty percent of

the cost of success of the hardware likely
a person infectious stuff that and sixty

percent of the cost of software and most
of this software is education specific so

it changes from one product to the next by
going in and using sort of a cell raw and

all the things we would like to
be able to reduce that to half so

we need to be able to go in and be much
better doing this so that's one of the big

industry drivers between doing sort
of collaborative robots in this and

we're seeing a movement towards doing
digital manufacturing How can I go so

I can say that my table and

simulate model the entire process and
I press a button and it generates code and

it actually works so I know lots of you
that sit at your desk and generates code

but I know very few of you where it
actually works off the arms so you know so

so there we're really seeing a big process
of how can we do sort of this bringing

together all of these tools to be able
to really deliver on the floor so

there's several big projects both in
the U.S. and elsewhere in the U.S.

there's a big protocol A.B.M.
and I have to be a call to make

flat out of Vanderbilt where
the idea is could we build a car and

then automatic you press the button and
the entire factory and all the code would

get automatically generated we're getting
pretty close so there are some really

interesting opportunities
to be able to do this so

some of the developers are really how do
we take on the monuments how do we make

flexible mechanisms with very
high precision how can I take.

A floppy universal robot and have it
have an accuracy of of thousands of

a millimeter or something like that
that opens up interesting option to say

how can we do in process inspections so
that we don't think about just kinematics

structures but we think about intelligent
structures and there's a lot of

these other things that's really happening
that's opening up but also we're seeing

an interesting opportunity of
Saddam to think about rogue.

Butts and other things as a service so
imagine if I could buy

you know so I would tell you I own
a Porsche What if I could buy it so

the most days I'm just trying to drive to
work it could be a really shitty car so

you know let's assume that most weekdays
I would have a car and I could buy it so

it has ten horse powers service then
on the weekends I want to go and

have fun then I would invest in say the
weekend I want to have one hundred seventy

horse power in my car I want to go out and
have some fun during going to and

from work as long as it sort of
gets me about the hill I'm happy so

imagine we could buy a cars or
robot for that matter where you pay for

the kind of performance you would have and
based on this you would be able to we do

this is some really interesting
opportunities coming out of this anyway

I want to talk a little bit of also
about saw the research that's going on.

In my research group and I wanted to pick
out sort of two examples one of them for

doing sort of systems modeling for

somebody on the other one talking about
some of the stuff we do in guided robotics

so but before I do this I want to
talk briefly about the road map so

you all hopefully know that we
wrote a road map two thousand and

nine and I've revised in two thousand and
thirteen the sort of model and

what do we think interesting projects
are for those of you that a Ph D.

student you should think about
this as a catalog of P.S.T.

topics you can go there and there are
things we think you can solve five years

out ten years out and fifteen years out if
you're investors you go for the fifteen

year problems if you're not so ambitious
you go for the five year problems but

there is I think two hundred forty
different Ph D. topics in there so

it's an interesting sort of thing to
look at if we look at it from the from

a manufacturing perspective for
a lot scale manufacturing it's all about

cost at the end of the day if you
don't save them money they don't care.

You might say I'm the most fancy algorithm
in the world they're like good luck So

think about this it's about Angela
too it's about process speed.

For the for the small manufacturing
company it's about agility and again it's

about cost them it's about an agility How
can you build very flexible systems for

this so what we did was that
we design sort of here saying

what are the challenges for
the company is or what are the drivers and

at the end of the day if you don't
make money we don't care then

the gaps what's heart so it's robot call
collaboration how can we build systems

that are much more collaborative than they
are today how can we have robots that have

sensors in the out a feedback loop
most industrial robots have and

in a feedback loop with joint controllers
but no sort of overall control

how can you how can you do this how going
to do plug and play integration how would

you be able to do much more
flexibility today it's true that if

you buy no matter what robot you buy you
getting locked into a proprietary system

a kook a world that's the orange world
Fanuc Well that's the yellow world or

and maybe well that's the white world and
it would be nice if you

could actually interchange these that
sort of what we've tried to see with Ross

that it would be nice if we get
much higher interoperability

problem with soccer with process that.

Changes between different generations so

it sucks sometimes of nobody does
professional software engineering and

the quality control is not there but
we've got quality control and we've got

real architectures this would actually be
fantastic that's really what we're looking

for here how can we do flexible
programming how can we do and effectors

I would still claim that grasping is one
of our largest unsolved problems for

this and of course how can we get high
performance manipulators high performs

manipulators with something something
that could do ten plus quests a second

that's a high performance manipulator and
of course high speed mobile platforms so

that leads them to a number of R. and D.
issues that can be solved from Asia to.

Control systems and

anything in between so so there are some
really interesting I would turn to is out.

Here to do that is the total points to a
dead we can do interesting things here and

we can all do and that's probably been
driving but what we're trying to do so

I want to talk a little bit
about using system L. for

doing assembly which is work that was
started in my group with Huckabee and

then later on right now it's also being
done by Kim on leave from my group so

it's really about how do we
get to a high level language

how can we get to a language that's
a little bit more interesting to do.

If you use an industrial robot you
typically have to be an expert in Carell

or wrap a door or one of these which is
completely useless you know you should

never doubt like you know if you come on
and say I'm an expert in C. plus plus

I will say that's completely useless five
years from now it's going to be D. plus

plus or something like that so who cares
so the question is how can I get back to

really thinking about having a high level
description of what I'm trying to do so

we've been doing this primarily to be able
to move away from doing this simple things

one of the tools for
doing this is system now so

I'm sure a lot of you have heard about
formal languages like you M.L. so

unified modeling language the problem
with your male side doesn't really

model physical processes
Sissonville allows us to go in and

also model a physical process so
we have his contacts and like forces and

control all of these things where we have
real fiscal interaction with the world so

there is this and it's basically
largely an add on to you M.L.

to be able to do this you will find
that just lots of industrial companies

that are actually using system L.
to specify their processes so

what we would like to do is we would
like to be able to have system out so

that we can add qualifications to this so

I can verify am I doing the right thing
if you get a special occasion some

from somewhere if I specify I
want to be able to do welding or

if I was based I wanted to paint up
I would like to be able to know you

know is it actually doing what I
thought it was going to be doing and

the other thing is can I
also use this to be able to.

Very far I am I building what the customer
asked me to do so we would like to

be able to use these tools to do this
to do education and then using this and

going out something we started with Andrea
and we continue later on is to can we come

up with a relatively high level language
for how we would be able to do assembly so

moving away from the C. plus plus and
moving away from the basic primitives and

thinking about you know over here I have
for instance a perception group that has

to do with what do I need to know about
the wilds we have some movement actions

we have some constraints so can I say
I want to move and I want to pick this

up with a certain amount of force so that
I can still get my coke afterwards and

this is a really bad way of grasping a
piece of coke if you want to drink from it

so can I add those kinds of constraints
on to this so that I can actually

come up with a reasonable level of
description that would be generic when I

implemented on a kook a robot on the kind
of grow but or whatever and can I make it

relatively generic with respect to
the process I'm trying to automate so

we worked extensively with Boeing they
don't allow us to have real Boeing parts

in the lab so we had to go for
the toy version of the airplane.

They're very sensitive about this
you know I'm not a US citizen so

they don't really want to give me access
to stuff like this that's OK with us and

also makes a little bit simpler and
more manageable to be able to do this

the question is can I do this but can I
use the same level of description to also

be able to build cars so that I can so
that I could do this both ways so

I really want to understand what is
the appropriate set of skills for

this what is your probably a set of
parameters so that I can easily go from

one of these to the other and can I at
the same time do verification to guarantee

that it actually does what I wanted it to
be able to do at the end of the day so

can we now from this generate
a description for the respective ones and

then automatically come Pylos into
a description that runs on the on the real

thing so we can come up with these
descriptions of what is in alt we can.

Specify what are the processes
that are involved and

basically I can now go from my dear I
can use our skill level description to

then come up with a low level description
of what are the set of actions and

then think about this as being
the equivalent of Java so I now have

a virtual machine running on my robot that
allows me to automatically take distant

description transform it into something
that can actually run on the robot so

this would make me have a description
that if you learned this once

you would be able to basically
whatever you're going to do later on

it would be relatively simple to do the
other thing that he would like to have if

you like to have a platform abstraction
you would like to have an abstraction so

that whether you're running it on a cook
or a robot or a universal robot it should

be agnostic I don't want to be build
into a yellow world on organs wild or

white Well I just want to
build robots who will bots So

the question is can I can I
go from Can I build sort of

a virtual machine that would allow us
to do this so that I can basically

implement each of these fundamental
actions once on my robot and

once I've implemented once it would
be able to run for whatever the task

description is that you would like to
be able to do to build your robot so

to do this that one of the challenges you
have is that this still looks tedious

How do I do this so the way we've been
working on this is basically can we use

P.D.L. as a way of automatically
generating the detail description

to do this we need to have an eight and
initial state a goal state and

a domain description we can run it through
your standard planning language and then

it will actually come up with a plan to be
able to do this so here is an example of

where we use this description of being
able to assemble a structure and

then it comes up very quickly with
a description and here is an example of.

A movie where we've actually done this and
last year for our Boeing to.

We replaced the Kuka robot with the
universal robot that was the only thing we

did and we base it recompile and then it
ran again so the fact that you made it

a platform agnostic for big companies like
Boeing this is a big deal you don't want

to get locked into a particular paradigm
where you can only do one thing at a time.

Here it's actually a simple description
because we're doing it purely for it but

it's an interesting example of what you
can do by using these high level languages

to come in with this.

So it's really about how can we have sort
of a low level robot I don't care what

the robot is you for a robot at me I'm
too I'll get another robot next year and

then we just want to be able to implement
this set of skills on this robot and

then go and be able to do this the other
is that now I can basically have

one way and
I can go to anyone I'm Andrea students and

say can you teach this language
to one of your robots and then

whatever I implemented on afterwards it
should be the same that would be nice so

that's really what we're trying to do here
and we've tried to do this some simple

learning about immigration we haven't
tried gestures yet you can use P D L and

came on has been working on how would
you be able to do this on a tablet

again because most of your users that
are coming in the twenty year old

person that's coming in is trying
to get a job on the factory floor

has probably used them i Phone probably
use the tablet if they can use

that interface modality they will feel at
home so that's very important to be able

to do this so developing I want to talk
a lot about is the stuff that we do in

vision guided robotics which is really
about how can we get I mentioned early on

most of the robots have standard control
in the joints but they have no overall

sort of close to control if we closed
the loop in the out of control loop would

we be able to get much high accuracy then
we have today so that's what I'm after and

then the good news is that there
are lots of cat models out there

we can use those to do
this automatically so.

We really want to get from you
know how can we get from these

unstructured wells to
highly unstructured wells.

Most people's is how the home is fairly On
Star is fairly structured if you have kids

go home lie on the floor look over and
you will see rough terrain so

you know there is there is an interesting
but you know how actually built this for

doing relatively complex operations so
we really want sort of to

move from the control settings
to to an unstructured setting so

here is some of the stuff that.

John you're in charge was doing while he
was here so the idea is really can we take

can we take a picture can we have a.

Can we have a database of objects here so
think about them as being canonical

of use doing them in some sense
doing object recognition today or

doing feature matching is inexpensive
memory costs nothing so if you need

a terabyte of memory and we can give you
a terabyte of memory not a big deal and

using this we can then basically match
against any one of these get on a problem

and an approximate post as cement that
would allow us to do sort of relatively

feature matching instantiate a model that
might be accurate to let's say plus minus

five degrees or plus minus ten degrees I
can now take that model I can take and

it's map of my image I can basically
do feature math and figure out

what the differences find that error do
post estimation and then we can do real

time tracking So this is relatively
simple to do if we can come up with this

process we should be able to do mission
we should be able to do tracking and

we should be able to do really well so the
obvious question is how do I come up with

the models the good news is that there
are a lot of bored people out there

you know and they so if you go to Ikea and

buy furniture there is probably a freebie
model in Google's for you where.

House which is pretty scary you know I
wonder if people could do all those things

but you know there are lots of these out
there so so you can go for most things and

get afraid a model the there's good news
here is there's bad news here they're

going to resist there's tons of models out
there the bad news is most of these models

are decide for doing manufacturing they're
not designed for using information so

so what what John Young did with
some of his work is can I find a way

of automatically simplifying these models
down here so that they are suited for

doing visual perception and basically we
have a technique where you will take away

all of the stuff that doesn't really make
sense from a visual process point of view

so now we get to these highly simplified
geometric models that we can then use for

being able to do automatic tracking and
if we gave the MacA few seconds to

to do this then you can actually see here
are some good examples of how we can do

relatively good tracking.

So you can see by doing multi-process
baby doing a particle filter you can

run it is and it's very robust You can see
over here by having a reasonable motion

model you can actually do good tracking
and you can see that it does pretty good

instantiation in tracking and this is code
that we've now released in open source so

anybody can use it and they can download
these models and they can use it for being

able to do tracking and it will do fairly
good tracking even down to something like

less than half of the object and you can
see it does really good tracking here so

it's nice to be able to have this and then
I sing is that we can use this to close

the loop so that we can do for
reading modeling and.

And this and of course if you look
at it will say OK that's nice but

the pretty boring objects.

This is an example from work
this is sort of the most.

This is the most reliable tracking we have
where you can see the does very little it

is compared to all the clutter in the
background and we would still be able to

track this very accurately simply
because because you use a model and

if you look at the number of it is on
the object compared to the cloud or

allows us to do very robust figure ground
segmentation and still do tracking so

that we can do this in real time so
this will run in real time on

a computer with a with a with
a reasonable graphics card and

we have a two day version and
we have a free day version of this.

I mentioned earlier on you can go to Ikea
and you can buy all of these objects and

they would actually be very nice so so one

of the things that John Young tried was
to say OK let's test the hypothesis and

say Could I download Ikea
models go to a regular home and

actually track objects so
he used a standard connect scanner and

now you can actually see he's
using a particle filter and

you can see it actually does
pretty good tracking here so

if you're going to try and interact with
things in the aware home or somewhere else

you can more or less download most
of these models from google for

you where house go in track AIDS and
be able to do either navigation or

many place them in the those
kinds of settings.

Which is highly satisfied that we can
actually go in and do this really well

again this run so if you run it on a good
graphics board you can you can get

down to multiple frames a second you
can get to pipe in real time yet.

But so you can see it's a little
bit jerky when you run and

try to run it in at full frame rate but
it's pretty good it's good enough that you

would be able to do navigation you would
be able to go in and manipulated and then

once you get close enough you recognized
a object you would be able to go in and

use the today's camera to be able
to do interaction with this.

The other thing you can do this for sure.

Do this we're doing been picking
which is very interesting for

a lot of for a lot of examples that
you would want to do in industry so

using a point cloud cam ARE BEING
ABLE TO GO IN recognize these objects

being able to estimate the post being able
to sort them so that you can sort them out

here it's basically doing by we do
an extensive amount of offline learning so

that you basically process it to come
up with where is the discriminative

set of features you extract those
you put them into a has table and

then you can maybe use the HAs
table to be able to do recognition

doing sort of the full processing
pipeline online is really hard to do but

by having the A lot of the sort of
heavy lifting done offline during your

training face you can go in and you can do
this very accurately so here's an example

of a point out that comes out and it will
allow you to go in recognize these objects

in Norreys with significant inclusion and
with clutter and actually get

very good recognition results good
enough recognition results that you can

go in and you can do sorting of this so
so there's lots of interest in this

from an industrial point of view and
being able to go in and

do this you know the example we've used
a vision for is to be able to go in and

do process to mation this is an old
saying that Andrew and I did with.

And looking at how can we actually go and
do training do some simple interaction

here is an example of a process we did
about a year ago with Boeing of using one

of these launch robots which
Alex Lambert is now working on can we

do sort of model based tracking of this so
that we can go in close the loop and

be able to do this with an accuracy that's
better than a tenth of a millimeter So

we just have a new this is over
from the lap we have on tech way

where we have three of these last robots
and we just got a new K. off five hundred

that's getting installed so this is one
of those robots that we use as a role.

A coaster by Kuka I don't
think I can get I.R.B.

approval to set it up as a roller coaster
but that would be another fun application

for this so what we have these where we're
going to use them to close the loop and

the objective is by using the visual
processing we can increase the accuracy of

this robot by default will have
an accurate it's a round point

four millimeters for drilling airplanes
we need to get this significantly below

zero point one millimeter using mission
in the outer loop we think we can do this

in real time and so we have to
demonstrate this about a year from now.

So I'll wrap it up here and say what
we're seeing is that by using robots

to do this we can get a much higher degree
of flexibility than we've seen before and

by being very carefully about
how you model it how you

sort of structure your control process and
your perception process we can go in and

get a much higher accuracy
than you happen for.

But for us it's been very important
I really need to think about that

as the system solution came to solve
the mission part of the control part

of the language part you really need
to think about this as an integrated

system to come up with a way of doing
sort of the optimal deployment and

by thinking about how can we do these
high a level languages like system L.

We can model the entire process
sort of into the end and

that allows us to have sort of performance
guarantees on is it really going to do

the thing that we thought it was going to
be and for that it's very important to

really enjoyed all of these aspects and
with that I'll say thank you and.

I.
Hope you enjoyed your launch.

Questions Yeah.

So so it's I think you will
see lots of processes where

you would want to have a much higher T.V.
I don't think we will see any time soon

the lights out factory where
there are no people there I think

will continue to see people being
involved at some level of supervision and

some level of operation but I think it's
right you will not necessarily have them

standing on the floor next to the machine
so you'll have to be able to do tell

operation where you can go in and do
things and I think again we sort of come

back to how can we do this with a high
level language so you don't want to and

I don't want to do joint control of a
robot to do this I want to be able to say

go pick up the Coke bottles of my
building in the perception and

intelligence you can operate at a level
what we are actually confident doing this

most row you know so
the best example I have is sort of

Wayne stuff what he's been doing on
the John Deere if you look at the back hoe

people are doing joint control of most of
the backhoes today people are not the sign

to do joint level control of any kind
of actory do as we should be able to do

in Defect or control in sort of Cartesian
space rather than in joint space so

there I think we will
use the perception and

the control to get us to a level that's
appropriate for people and then I think we

can take out some of the delays and other
things to also be a good accessible so

I can sit here and
operate a backhoe in Australia.

OK.

Enjoy the rest of your day.