So, a large part of what you’re going to see

here was worked on at UT Dallas, just want to give that little disclaimer but we’re very

excited to be here at Michigan with electrical engineering and the Robotics Institute, and

so we’ll be doing a lot of cool stuff here over the coming years. So the motivation behind

my talk is challenges to human mobility. And in particular, we’re very interested in

challenges faced by lower limb amputees, so there’s about a million Americans with lower

limb loss as well as general mobility limitations that can be caused by stroke, advanced age,

Osteoarthritis, lower back pain, musculoskeletal disorders, et cetera et cetera. In fact, it’s

pretty staggering numbers about one in eight adults in the United States have some sort

of mobility disability. And so, the features of gait experienced by these individuals include

more frequent falling, more metabolic energy consumption walking from point A to point

B. And slower walking and so forth. And so there’s certainly a lot of room for

improvement of conventional prosthetic and orthotic devices to enable greater mobility

in these populations. And so our approach, my group likes to think of legged robots as

a cycle of life for achieving these goals and so we like to borrow and develop concepts

in the area of legged robots, autonomous robots like Cassie and translate those into powered

prosthetic legs, learning how to cooperate and synchronize to the human with the robotic

device as well as in exoskeletons trying to augment an existing limb that may have some

sort of weakness or deficit to it. And then through the process of learning how to apply

these control methods to humans we actually learn a little bit more about how to make

more human like autonomous walking as well. So we can make these robots walk in a more

natural lifelike manner. So that’s the cycle that we embrace. And so the organization of

this talk is as follows. I’m going to start by discussing our efforts in user synchronized

kinematic control of agile powered prosthesis. And so when I say kinematic control I mean

controlling joint position and velocity. Then we’ll talk about energy aware actuator design

in order to enable greater mobility in these devices. And then we’ll close with discussions

on energetic design and control for partially assisted exoskeletons that only provide some

support for an individual, not complete support. And so you’ll see how we’re kind of going

from this kinematic control paradigm towards this energetic control paradigm in this talk. So this is the state of the art in the field

of powered prosthetic legs. At least it was when I entered the field. And so the idea

is you look at the gait cycle and you notice it has several different phases of gaits which

are indicative of certain behaviors such as heel contact, push off during late stance

and then the swing phases. And so essentially the way that control engineers were designing

the powered devices for amputees in the past is to design a different controller for each

of these behaviors. So each of these boxes may have a proportional

derivative controller in it. Sometimes in the field we call it an impedance controller.

And then there would be some switching rules to go from one mode to the next. Now the problem

is that all of these things are subject dependent. So I walk differently than Rom, Rom walks

different than Gray, and so on. And so we actually end up having an explosion of parameters

that require patient specific tuning that is not intuitive because these things may

not necessarily have physical intuition behind them, especially the switching rules and some

of the parameters depending on which control methods you’re using, and it takes a lot of

time. And the problem gets even worse when you think

about variable activity locomotion. So typically you’d have a higher level state machine where

you’d have let’s say five different modalities you can use and each one of those things would

have its own lower level cycle here with its own parameters and so what ended up happening

is you’d spend, and I’ve done this during my post doctoral training, you’d spend hours,

sometimes days with one person trying to optimize these parameters to make them walk just right. And so it’s not a clinically viable solution,

right? Medicare, insurance, whatever is not going to pay for five hours of tuning with

a team of engineers. And it’s also not necessarily very accessible to people who haven’t had

training in control theory such as clinicians. And so that’s the challenge. And so I took

inspiration from the field of robot walking to address this challenge. Now, this is a

robot that was designed by Eric Westervelt who I believe is a Michigan alum, right? And

then later adopted by Jim Schmeideler at Notre Dame. And so Ernie, has he’s called, you’ll see

in this video is controlling its hip and its knee angles as a function of its hip progression,

okay? So it’s measuring the progression of the center of mass of the hip. And so as the

robot’s pushed forward, its leg joints will track a pattern in the forward direction and

when it’s pushed backwards it tracks back in a backward direction. And then when you give it a nice push it can

actually fall into a nice self-fulfilling walking cycle were actuation energy injection

from the leg joins cause more forward propulsion which then keeps the joints moving forward

and so on. Now this idea actually initiated from Jessy Grizzle here at Michigan. And this

is the notion of a phase variable, okay? So it’s a time invariant parametrization of the

joint trajectories. So let’s say that you want to track this kinematic pattern, so this

is joint angle over time. But you want to track it in a time invariant manner so that

you can, so that the robot can always keep up with its progression, right? So that the

leg is never too far behind or too far ahead. And so what you do is you try to find, at

least during steady walking a monotonically increasing signal that you can measure, which

we call a phase variable. So this could be the hip position going forward

in space over time. And so then if you measure that phase variable you can then parametrize

your kinematics as a function of that. And the beauty of this is that if you’re walking

and then suddenly someone grabs you and stops you from moving, then your joint angles which

are being controlled as a function of this phase variable will stop, right? Synchronizing

with your progression and then as soon as you let go of the robot then it will keep

moving again, right? And so you end up seeing this … so you see

this little phase shift here that’s happening over time but because we’ve represented the

trajectories as a function of a phase variable it’s as if nothing happened, right? Because

you never left the desired profile, you just simply slowed it down and then restarted it.

Okay? So that’s the idea of the phase variable. And it’s turned out to work really well for

walking robots. These things can walk and run and climb stairs and so forth. You can

push them and they’ll recover. And that’s still going on here at Michigan with Jesse

and Rom’s groups. So we wanted to use this for controlling a

prothetic leg. But that then motivated the question of how can we measure a phase variable

that represents the progression of human walking? And so, my former PhD student, Dr. Villarreal

came up with this concept of the thigh phase angle which is a strong predictor of the distal

joint patterns. And so the idea here is that if you’re measuring the angle of the thigh

with respect to gravity, so this is a global and inertial measurement then you look at

it over time, it has a sinusoidal trajectory. And so when you have a sinusoidal trajectory

like this you can build a phase orbit. This orbit in the phase portrait actually gives

you a clock, a sense of timing which is actually time invariant because it’s based on the position

of the thigh moving over time. And so in this study, Dr. Villarreal even

showed that it’s, the distal joint patterns are still highly correlated with the progression

of the thigh even across perturbations. So if you tripe someone, at least with a reasonably

magnitude perturbation you can still predict distal joint patterns with this measurement.

Now if you really trip someone and you’re to the point of nearly falling over, then

there might be some reflexes that kick in that might not be well explained by this.

But at least subtle perturbations are well captured by this idea of a phase variable. And so we implemented this concept in our

first leg design, this is our gen one robotic leg. It wasn’t designed for looks, okay? So

it’s not the prettiest leg but it was designed, it was our first attempt at building a wearable

robot. So the key features here is that it has an inertial measurement unit at the top

of the knee joint which gives you the orientation of the thigh, okay? So it has an accelerometer

and gyro in there. And it has a knee actuator and an ankle actuator. And I’m pointing out

that it’s highly geared, I’m not trying to brag about that. I’m just pointing out that

it’s highly geared because that’s going to become a problem later and I’m going to talk

about how we’re going to solve that as well. Okay. Definitely I’m not bragging about that. And this was work by Dr. Quintero who is now

at SFSU. And so having this implementation in the prosthetic leg, we also need to figure

out how are we going to define the kinematic patterns that the leg will follow. And so

what we do is we start with able bodied, normative kinematics which we then reparametrize again

as a function of this phase variable. But then we allow the clinician to visually augment

that trajectory. So you can imagine that the patient puts on the robotic leg, starts walking

with it. But because every gait should be user specific, should be unique the clinician

may observe okay, they need a little bit more push off during late stance or maybe a little

bit more inflection during swing. And so then they can grab these control points

and manipulate the trajectory and then it re-encodes the function of the phase variable.

So essentially we have these functions of phi which is the phase variable and that then

leads you to an error vector where theta is the actual truly, the measured angles of the

leg. This is the desired angle of the leg, given the phase. And so you have an error

vector. And then the simplest way to control the leg is to control torque based on a proportional

derivative control law, right? That’s the first thing we all learn. And so we do that, it works. But there’s also

more rigorous formal methods such as hybrid zero dynamics and Dr. Martin has a really

nice paper on that theoretical approach but we’re not going to use the theoretical approach

in the experiments just because it’s difficult to model these things and it could result

in other challenges. So, and the experiments to follow were using partial derivative control. All right, so this was our very first amputee

subject. The very first time anyone had put on our, any patient had put on our leg and

this was just him acclimating him to it. And we were just recording the video while we

were shooting the breeze with this subject. And I highly recommend doing that, record

anything just in case something interesting happens, even if it’s an outtake, it could

be useful. And so here, we’re just chatting with him

and then we notice that he became kind of comfortable with the leg and started kind

of moving his weight, shifting it forwards and backwards. And then we asked him so what

are you doing here? And he was like well I understand that the foot’s going to be where

I expect it to be based on my hip motion so I can actually trust that it’s going to be

behind me when I need it to be. So he’s not looking at it, he can’t feel it,

right because he has no proprioception. But he has learned this mapping from his hip motion

to the prosthetic foot position. And so essentially he was able to then volitionally to a degree,

it’s not true volitional control but he’s at least able to control the foot position

where he wants it to be. And so that was pretty nifty. And this was working in collaboration

with Susan Kapp. And so here’s a sequence of experiments demonstrating

how walking with a conventional prosthesis which is what we’re going to see first. This

is the leg that this subject uses every day and then comparing that with walking with

our powered leg. He’s actually a pretty good walker with it. So it’s a little loud and

we’ll talk about that later. That has to do with the high gearing and the use of a ball

screw. Now walking backwards is hard because with

this conventional leg he has no control over foot position. But against with the powered

leg and this controller he can trust that the foot will be where it should be based

on his hip motion. So Bobby it looks like he’s sort of snapping

the leg into place as he’s swinging backwards? Yeah, that was because we have … I’ll talk

about it in a second, yeah. Yeah, but you can go forward and backwards. And we can also

do things like crossing obstacles which the controller was not explicitly designed for

that. But we were able to, based again on this mapping from hip motion to foot position.

Okay, so Rom, the question about the snapping. So we, I’m kind of trivializing things a bit

here. We have some safety mechanisms in place so that the leg doesn’t for example switch

from stance to swing erratically or when you don’t want it to do it. So in order for the

leg to start flexing, before you actually take a step forward we require it to have

a certain amount of motion to actually engage we call it backwards stated. But it’s all

based on the phase variable, though. It’s just that there’s some supervisory logic that

prevents it from switching too quickly. This one was a little more entertaining, so

we got a pretty good soccer kick out of this. Whoa, nice. So … better than I can do at least. All

right. So, and so here’s steady walking. This participant does keep his hands on the hand

rails but this might have to do with the actual design of the device. So trust me on that

one. So you get more normative energy injection

into the gait cycle. And so on particular, for example while we’re walking at multiple

speeds, normatively it would be the knee work gets more negative as you go faster because

your knee is doing more braking as you walk faster. And ankle work would go up because

your ankle is doing more propulsion as you go faster. And so we get that, whereas with

the passive conventional leg, this would be more or less level and it wouldn’t be nearly

as high, either because it’s a net zero mechanical work paradigm, right? I mean these prostheses

conventionally do have springiness in them so they can store and release energy but they

can’t inject energy. And so this results in more normative biomechanics

and that’s still a somewhat theoretical concept but this is where things really matter. So

we look at compensations of the amputee participant. And there are three that are very common in

amputees. So there’s hip circumduction where they kind of rotate their hops so that their

foot doesn’t drag the ground. There’s hip hiking which is exactly what it sounds like,

again to prevent the foot from dragging on the ground. And then on the sound side they

have ankle vaulting where they kind of push off too early to again provide ground clearance.

And what we see compared to the passive, so the dotted line is the passive and the powered

is the blue. We see actually a reduction in these compensations. They’re not completely eliminated and we wouldn’t

expect that after one experiment but we do see an almost immediate reduction in them.

So this is relevant because these compensations are metabolically costly. So also they tend

to wear the joints, the intact joints fast. So amputees tend to have arthritic hips and

arthritic knees in the sound side and so forth because they are overusing their sound limbs. Okay, so at this point, I described a continuous

sense of phase, but we still have a discrete sense of task in the sense that we would have

to know what the task is, the activity in order to change the kinematic pattern of the

leg. Right? Because so far we’ve just made the kinematic pattern time and variant but

there’s still a profile there that we’re tracking. And so in order to do different things like

walking on inclines or stairs, you’d have to change those kinematic profiles somehow.

And so you could do that with a classifier. But my goal and the topic of our current R1

project is to have a continuous sense of activity, of task in order to allow navigation over

continuously varying speeds, inclines and other types of transitions to stairs and so

forth. And so the goal here is that we can define

a multidimensional activity space where for example, slope, the incline can be any real

number within some range, speed, can be any positive real number within some range. And

then for example stairs. Are you on a staircase? One. Are you not? Zero. Or are you transitioning

between them and that would be some number in between zero and one. And then our goal

is to have a kinematic model that given a point in this activity space will then give

us a prediction of what the kinematics should look like. Okay? And so you see here that for these three different

points in the activity space we have three different trajectories that we’d want to track

in the prosthesis. So that’s the goal. And so we have some preliminary work in this direction

where we’re trying to use samples of these … of this activity space. And so in this

study, our activity space is just variable inclines and speeds. And so we do a motion

capture study where you can, where we sample a certain number of these speeds and a certain

number of these inclines and combinations of the two. But we can’t possibly sample the

continuous range, right? That’s not possible. And so we need a model that can predict in

between those samples. And so in order to avoid over fitting, we

wish to use regularization to find a model that explains the underlying features in the

data in a way that can allow us to predict activities that you don’t actually have samples

of from the motion capture study. Okay? And so for example, this function B here would

be a function of the phase variable, okay? And that would be, for example that would

come from a set of trigonometric polynomials, and C would be a function of a vector of describing

the inclination, and speed I guess. I projected away speed. But speed would also be in there

and so essentially we use group L1 regularization, which is a machine learning method for trying

to induce sparsity in the model. Okay? And so we’re able to show that this is actually

a stronger predictor of unknown, of untrained activities compared to for example linear

interpolation which is exactly what we would all try first, right? So essentially our error,

our prediction error is statistically significantly better using our sparse model than just using

linear interpolation. And so the argument here is that then we can

reliably predict the desired kinematics for any incline or speed that we can detect from

the environment. Okay. Now detecting the incline and speed is a challenge in its own right,

which we’re working on. But this is the idea. And so, we’re currently working on modeling

and control of again, these continuously varying activities. But we also want to consider stairs

and maybe one day running. The ultimate goal of course having a control system that can

perform all of the activities of daily living in a seamless manner. And so we have a way of doing that theoretically

at least in terms of transitions between a certain set of activity modes, like stairs,

flat ground, sitting. And then each of those modes might be parametrized by an incline

and a speed that would then allow us to have this continuously varying set of activities.

So this is an early project that’s still ongoing. However, in order to actually achieve this,

we’re going to have to address the limitations in the hardware that some of you have already

pointed out. The hardware is loud, the hardware is heavy, it’s got umbilical cords attached

to it. And it’s very stiff. When we use these highly geared transmissions, it makes the

joints very stiff in the sense that they’re not back drivable. You can’t … they don’t

naturally swing with dynamics and so the motor has to literally do everything for the robot

to move. And that’s not how human joints work, right? And so this is where now we’re looking at

using better hardware such as prosthetic legs that actually think about how they use energy.

In particular, the open source leg was designed by Elliot Rouse here at Michigan and we’re

one of the lucky early recipients of it. And it has a series elastic actuator. So essentially

you put a spring in series between the gear box and the load which provides compliance

to impacts. It also allows you to store energy and release it. And so it has the potential

to reduce energy consumption in that manner. It also provides some back drivability and

lots of different potential benefits. But if we’re going to be using this for controlling

variable activities, that begs the question of how should we select stiffness for the

series elastic actuator? And so in Elliot’s design he has six different

selectable stiffness options you can choose, but you can’t pull them out mid gait cycle,

right? You can’t change the stiffness as the leg is being used. And so, at least not yet.

And so the question is how do we select an optimal stiffness that will allow energy efficient

electrical energy consumption as well as satisfying actuator constraints? For example you don’t

want to bottom out the spring, because then it becomes a rigid actuator as soon as the

spring bottoms out. And also limitations like the peak torque

and velocity of the motor and so forth. And so that motivated a different NSF project

which is in its early stages where we’re trying to have a method for robust design of series

elastic actuators. And so it turns out Edgar, Dr. Boulevard is here in the audience somewhere.

There were go. So was a PhD student with me and now a post-doc with me. So Edgar realized

that you can express the energy consumed by the motor as a quadratic function of the spring

compliance. So that’s the inverse of stiffness, right? And so, what is energy consumption? Well it’s

not the energy to move the load because that can’t be reduced. That’s just first principals.

But it includes energy losses due to a viscous friction. Okay, so friction results in loss.

And joule heating. So that’s the heating that comes from Joule’s law, the windings in the

motor. You put current through it and it heats up with, scales with current squared. And so we can potentially reduce those things

through the design of the spring. And so, in Edgar’s analysis he was able to show that

you can actually express this as a quadratic function, energy consumption. And in the case

of a linear spring, meaning a constant stiffness, that means that you have a convex function

of energy over compliance. So you can very easily find the optimum that

minimizes energy consumption, right? Now if you’re thinking about a non-linear spring,

well then X would be a trajectory of compliance that has certain constraints on it. For example,

you want your spring to have a monotonic relationship between displacement and torque, right? So

in order for it to be conservative. So we have inequalities that correspond to actuator

constraints and feasibility constraints. And we can also introduce uncertainty. Because

this is a convex problem, a convex optimization problem there are tools available for robust

optimization that can handle uncertainty. For example, our uncertainty in the case of

a legged robot would be for example the mass of the human user. We don’t really know how

much each person weighs or if they’re wearing a backpack. You put a bunch of books in there

to go to class or whatever. Or iPads these days, right? And the position might be a little bit uncertain

right? Because the environment might result in differences in kinematics and so forth.

The efficiency of the actuator itself might be uncertain and there could be all sorts

of unmodeled dynamics as well. So now this allows us to again minimize energy consumption

and also guarantee that we still satisfy actuator constraints as the activity might change. Now there’s another approach that we’ve been

investigating which is not using series elasticity but instead using quasi direct drive actuators.

This is a different way of achieving compliance, it’s just that it’s not compliance through

a spring, it’s compliance through a lack of inertia in the actuator. So the actuator does

not have significant dynamics of its own. So you can just almost freely rotate the motor

from the load. It’s called back driving, all right? So the way we do this is we have to

have a very low gear ratio because the inertia reflected, the inertia of the rotor, of the

motor reflected through the transmission scales with the square of the gear ratio. And so the whole game comes down to getting

this gear ratio down as small as possible in order to reduce the inertia of the actuator.

But then when you reduce the gear ratio then you have to deal with making sure you have

sufficient torque coming out of the actuator as well. And so that’s where the high torque

motors come in. We have to use high torque motors that typically are pancake motors because

they have higher torque density. And these two things combined allow us to do some pretty

nifty things with our Gen-2 leg. So, you can see it’s very back drivable. This

is powered off. It requires one to three newton meters of torque to back drive the knee or

the ankle, so it’s very back drivable. And the goal there is to allow more dynamic motion

and also energy harvesting. Because when this leg is doing negative work, when its braking,

the motor is doing that. And that results in a charge going to the battery to prolong

battery life and so forth. Also, there’s fewer moving, meshing parts and those parts are

moving slower than a highly geared transmission which results in less noise. So we can finally

address that problem with the lawn mower sounds coming from the prosthetic leg. All right, so here’s the video I promised

you where the subject will wake without the hand rails at some point. So here it has very

compliant impacts with the ground because there’s almost no inertia at the joint. Of

course the limb itself has inertia. And it’s also able to have a very fast push off to

swing transition because it has a very high bandwidth. It’s one of the benefits of a quasi

direct drive actuator is very high bandwidth. So you can go from high force at push off

to high velocity at early swing in almost no time. The battery is right here. Yeah.

And it’s enough for about I believe, it depends on walking speed but it’s enough for around

5000 steps or more. Oh, 50 decibels adjusted is about the noise level of a household refrigerator,

so very quiet. Something that at least we find acceptable at home, right? And again

we’re not stuck to treadmill walking. And one of the other benefits of the quasi

direct drive actuator is again I mentioned earlier energy regeneration and energy sharing

between the joints. So when the knee is doing negative work and the ankle is doing positive

work they’ll share that energy rather than the battery having to provide all of it. And

that resulted in a specific power of about 50% less specific power than the state of

the art Vanderbilt prosthetic leg. So we’ve cut energy consumption in about a half. So in the last part of this talk I want to

switch topics to exoskeletons. However, still very related to the latest leg I showed you

in the sense of its actuation paradigm. And so the state of the art for exoskeletons is

they’re primarily designed for spinal cord injury applications where the human can do

very little to move their own limbs. And so the exoskeleton has to do everything. And

so in this context it makes sense to design very stiff actuators so that the weight of

the human subject doesn’t drive the actuator, the joints so they don’t collapse, right? However, that stiffness means that these are

not very useful for working with stroke or anyone who has voluntary control of their

limbs. And so, where my lab is going with this is going towards a more partial assistive

paradigm where instead of having stiff actuators, rigid actuators we have back drivable actuators

and we use quasi direct drive designs to achieve that. And also, we don’t want to use kinematic

control in the context of partial assistance because again even with a back drivable actuator,

you need the controller to not be creating forces that oppose the human’s intention,

right? So if you’re controlling kinematics then the

human must still follow the kinematic trajectory that the robot tells them to follow. And so

that’s why we’re heading towards energetic control objectives. And so our patient populations

of interest are stroke, OA, advanced age and overuse injuries. So as I hinted at we have

this quasi direct drive paradigm where we have a 24 to one gear ratio in this particular

exoskeleton which has a powered knee and a powered ankle. Yet, we’re still able to produce

large torques at each actuator, so about half of what they would need in everyday life.

So this is actually still a quite powerful exoskeleton in terms of partial assistance.

But only about one newton meter of back drive torque. And it appeared on the cover of IEEE

Control Systems Magazine a couple of years ago. And so in terms of how we control it, we are

using a method called energy shaping, otherwise known as Lagrangian or Hamiltonian shaping.

And the idea here is that we model the human body as a Lagrangian. Okay, the Lagrangian

is a scalar function that is kinetic energy minus potential energy. And the reason that matters is because if

you put it into the Euler-Lagrange equations it spits out the equations of motion, which

we see here. So the m for example is the matrix of masses and inertias and how they depend

on joint angles. C is the matrix of Coriolis centripetal terms and g is the vector or gravitational

terms. And we have orthosis torques which of course

influence the dynamics. And so in energy shaping, at least the way we’re using it, we are designing

the control input, u, such that when you close the control loop, the close loop system dynamics

behave like a different mechanical system, okay? So now this mechanical system corresponds

to a different Lagrangian where mass and inertia have been reduced. That’s the idea. We use the control torques at the orthosis

joints to reduce the perceived mass gravity inertia of the human body so that they can

use less muscle effort to move their limbs and to fight gravity and so forth. So this

is the idea. And in the case of under actuation which we do deal with here, this is challenging.

There’s something called the matching conditions that need to be satisfied to show that there

exists a control law that gives you this closed loop system. And that’s not trivial, it’s

a set of partial differential equations. But we have some clever ways to deal with

that and Jinping is in the room somewhere and working on that hard. And so here’s a

demonstration of our knee and ankle exoskeleton doing energy shaping on the human user. And

you can see that he’s able to sit, stand, walk freely. Now this is an able bodied user,

so he could do this normally. But the fact that the exoskeleton isn’t stopping him says

a lot. Okay, because it’s very back drival, he’s

able to continue to be in control of his movements while being supported by the exoskeleton.

And so we’ve done some analyses of this. We look at EMG activation. VM means vastus medialis

and we are able to show that with the assistance, the active mode, so blue is the EMG activity

when the exoskeleton is on and powered on we see that it’s lower than the case of bare,

without the exoskeleton and passive is wearing the exoskeleton but without the motors on.

So we actually see a pretty substantial drop in the EMG activity of this particular muscle.

And red is the torque showing that it’s doing something, to do that, right? So this is for sitting, sit to stand and we

see also promising results for walking. Where again we see in blue, this is the EMG activity

of the soleus muscle which is reduced during, especially during push off with the assistance.

Okay, so this is my last exoskeleton I want to talk about. And so this is a powered knee

only and it’s a much smaller scale exoskeleton than the one I showed you earlier. And so

this is meant to be a conservative treatment for OA and also to prevent lower back strains.

So for example when people have to lift things repetitively in warehouses, in the military,

assembly lines, et cetera. We want an exoskeleton that can assist that so that they don’t fatigue

and then use their back in an unsafe manner. And so here because we’re primarily targeting

individuals who are more able bodied, who have very minor impairments if any at all,

we took this to the next extreme where we have only a seven to one gear ratio. And in

order to reduce the gear ratio that much we had to come up with a custom design for an

electric motor that can product higher torque for longer periods of time. And so in order

to do that we use encapsulated windings which have a more efficient heat transmission so

that the heat around the windings of the motor get distributed to the environment in a more

efficient manner. So that way we can have a higher continuous torque. So this is less than half a newton meter of

back drive torque and up to 20 … now I should update this. Now it’s up to 25, right Chris?

25 Newton meters because we improved the magnets. Okay, so this just demonstrates it’s very

easy to move around with this thing. And one of the applications is assisting stair ascent

because that’s a potential home use application, especially for advanced age. And he’s the

lifting and lowering experiment that Nikhil conducted where we have a … how heavy was

this? 20 pounds? Yes. 20 pounds. And we have a force plate to make

sure that the subject isn’t biasing one leg versus the other and then we’re recording

EMG activity of multiple muscles but we’re only going to show one of them. And so this is bare mode, this is just a baseline.

This is passive mode, so wearing the device but it’s powered off. You see very minimal

difference which is actually kind of good in itself, it’s very back drivable. And this

is active mode. And blue is the EMG, red is the torque coming from the exoskeleton. And

so you see that EMG has dropped. And this is the rectus femoris, yes. So. All right, so this a clear picture of this

plot. So you’re able to see again a reduction in the muscle activation. And the goal again

is to prevent fatigue so that proper lifting form is maintained for longer. So in closing

we have several ongoing studies that are more towards clinical outcomes so we’re looking

at functional outcomes for stroke subjects, assessing muscle tension and posture during

lifting, lowering and carrying and assessing muscle tension and pain in knee OA subjects.

We have enrolled one OA subject but the results, we’re still looking at the data. We haven’t

really processed it yet. And then a very quick overview of some other

projects and I’m sorry that I’m missing some people from this presentation but we have

several other collaborative projects back at UT Dallas as well as Virginia Tech and

UT Arlington. So in closing I just want to recognize the contributions of my lab members,

and this is just a subset of the lab. But this was the last photo we took in Dallas

before the big move and then the funding agencies NH, the NSF and the Burroughs Welcome fund.

So I’m happy to answer your questions.