It is a. Distinct pleasure for me to be introducing and welcoming her to this for you know what's good for us to do good days I'm really happy and all Irish. You know we're trying to leave. Off with all the audience strong. Feelings for the robot the boy systems of today we're going to be free to have Cyber a human interactions with a good first thing. Rather than the border a lot of the stuff that goes on here on campus who are. Not only looks at cyber human systems but the U.S. is really a not only a control guy at heart as far as I know meaning is higher has been the norm in your control think he's throughout my career we've been able being various incarnations all the water systems all the way from when he was a fifty two students he got a Ph D. from. A class a university and. He already cracked something small I feel guilty for years morning but. I think her birthday. That happily graduated went to all group stuff a lot and I think that he went to University of Florida going back to the data golden era phase the part where he's basically go through the ranks and he's now a and then doubtful professor in. This department and his work has been both practically relevant theoretically elegance is one of the few people that truly can struggle this finds in the fact that matter in his work has been highly recognized and rewarded correctly so so I got tired of reading his bio because they're worth every If he is a fellow told tales of the I simply this one the future. More. Yes a career of more science from the early career of the work for all that I see a lot of. US will still call the commanders of the service of war. US Air Force so that's where the salt really got going ordered field by. And in fact this is more of committees vote only with resources but from the. Always when I feel that I got to do that I look at it when I go yeah why it won't work because you see me is ridiculous. When he sleeps we don't know you know he never shapes nobody. This is a more rugged us a little bit more of we're delighted to have you there for sure thanks for having us it's a real pleasure to be here in fact today's talk. In some ways has its roots here in Atlanta when I worked at National Laboratory we were doing a lot of some of the early days exoskeleton work and. I imagined how exoskeletons could be used for rehabilitation and I took it from a very naive roboticists approach where hey robots can do things more repetitive and more precise than people can. And it's a lot of physical labor for physical therapist to move the limbs of people to do physical therapy so why can't we just replace the physical therapist for the robot and I visited the Shepherd Center here in Atlanta because it's one of the largest rehab centers in the Southeast and I was in mightn't how naive that impression of physical therapy was and that there will always be a person in the loop a physical therapist in the loop there because of the interactions with the person. And also with. The exoskeleton work at the time having a small portable power supply was really one of the limiting factors for having soldiers carry three hundred pound packs and sprint through the mountains of Afghanistan. And that was part of my motivation for looking at rehabilitation because you could be in a lab and just tethered to the wall. But was it a Shepherd Center I saw for the first time muscle stimulation and it instantly dawned on me while we are to have nature's power plant right there waiting for us to use. Its low power we eat you know just give us soldiers some to read O's and then they've got the fuel that they need to to actuate their muscles and then I saw the rehabilitation aspect of it and. From that experience here at the Shepherd Center in Atlanta I then wrote my career proposal almost topic and I've been working on it ever since so in some ways my trip to the Shepherd Center really inspired. The initial development in the Syria. I want to think my students first and foremost because really it's their creative ideas and hard labor I actually do get a lot of sleep as these guys who don't get to sleep very much. And I give a lot of credit to these guys we've had a number of formal muscle guys in the group who have gone on to faculty positions several of them have started their companies one of the companies has been they've received some multi-million dollar buyout offers and they've said no which I think is crazy great for them. And so it's been a real success for me or pleasure for me to kind of see people who work in this area go on to do great things as well and so this is my current group of students some of which work on this problem others of which work on some other problems. That's new. The different colored T. shirts don't mean anything it's just some people want to blue and some people want to black ones. I don't know what's happening with my second slide here seconds ago it was fine would move on past that it basically was just showing what is muscle stimulation and if you if you've watched late night T.V. And you've seen these as seen on T.V. ads where you apply stimulation to a muscle it'll contract. I don't know necessarily to give you the six pack abs it certainly I don't own one but I don't think it would have helped in my case anyway. But it is a commonly clinically prescribed. Rehabilitative treatment and. And that's exactly it you apply a potential across a muscle and it will contract and so this whole art of closed loop control of the muscle is how to intelligently stimulate the muscle so that actuates the limb and moves the limb in a desired way. There are a number of challenges associated with this from the controls community and while the clinical community has been doing muscle stimulation for one hundred years the controls community really hasn't embraced it until maybe the last twenty years and even there it's really been the adoption of controls tools without necessarily understanding the underpinnings of those tools in order to kind of use them maybe in the most optimal way or to develop them further and so whenever I give this talk I really it's really a call to arms for people from the control systems robotics background to maybe think about this is an application that application outlet for your research. So in the upper videos. See that the muscle contracts differently depending on what frequency worse we're stimulating it at this video seen on the left seems to stop playing but you probably saw it bumping around that was if you had a low frequency you could actually see the low frequency stimulation in the right approach most video this is a higher frequency stimulation was just a constant contraction we'll talk a little bit more about that later and then in this bottom picture which will also talk about later you see the geometry of the muscle changed quite a bit so where you may have electrodes placed and you stimulate when the arms at a certain angle at a different angle the muscle has moved and now you're going to get a different response from the muscle and so how can we compensate for that. Muscles are inherently nonlinear and dynamic systems they also have a large number of uncertainties associated with them my muscles and your muscles will be very different to have a different response to them. It's a function of your hydration of the person the fat content of the person different people have different muscle fiber types are different compositions of muscle fiber types some people are more inclined to sprinter muscles some people are more inclined to more marathon muscle. And so and the muscle fatigue is very rapidly when we externally stimulate it one of the questions that people often have is well are you stimulating it like the brain stimulates it and we can't because the nervous system innervates the muscle fibers in a way that we just don't have access to so we we don't have access to the individual muscle fibers to stimulate them in a way that would would be akin to the way the brain would do it and we don't know the brain knows when it stimulating. Marathon muscle fibers versus sprinter muscle fibers for example and we wouldn't have that information so when I talk to school kids you know describe this is we just put stickers on your legs and then re and. Enable your legs to do things and especially for kids who have spinal cord injuries or a variety of neurological disorders these stickers can and they will them to do things they couldn't before. And so based on the fact that. Muscle fatigue rapidly because of the way that we have to externally stimulate them the fact that there's the sense certainties the fact that there is these non-linearities. And other factors that we'll get into in a few moments these cause real fundamental challenges for us to examine from a controls perspective So some interesting questions that come up when thinking about this area more from a controls a control systems perspective some of the things that we've been thinking about in my lab are how do we develop an overall stable subsystem when we switch between modes where we're controlling a person and when we're not controlling a person or when we are controlling the system the stabilize the ball and a system that is unstable and we want to be unstable so so developing controllers for that is is a question. If we're switching between controlled systems and uncontrolled systems for those of you who are maybe familiar with hybrid systems theory and those that that means you have to develop as well time for how long you spend in the controlled system versus how long you spend in the uncontrolled system and to develop these these twelve times you have to know how fast you converge but what happens when you want to use adaptive control for example because adaptive control is really amenable to this kind of problem because we want to compensate for the uncertainties in the system but with adaptive control you typically are dealing with a non strictly optimal function and you don't know how fast you converge you're just converging asymptotically So how do you develop a time when you're only converging asymptotically. How can you and how can you compensate for. Uncertain input delays I'll talk about that problem in just a moment. And a recent thing that we've been thinking about is is as a human robot system two different systems that are interconnected in some way and if so then maybe passive a theory is the best way to look at those kind of problems where you want the robot to be passive with respect to the human for example or is the human robot system one continuous time Amec system that just has a logical switches between the two and in that case maybe hybrids which systems theory is the best theories to kind of approach the problem from and we're kind of looking at it from both perspectives and exploring the pluses and minuses of both of those and then related to the functional electrical stimulation question and functional stimulation is when you're applying electric field to do a functional Pasch like cycling or bicep curls or those kinds of things that you'll see in the presentation. And. Traditionally electrical stimulation was applied as a means to just build the muscle mass or to keep the muscle from atrophying you just don't want the muscle mass to decay because of disuse and that was fine you can use open loop stimulation trains and you can put the muscle under load and you can do that but if you want coordinated limb motions to perform functional tasks and modern rehabilitation theory has shown that if you can do a repetitive task with some cueing that that helps the neural system. Reestablish itself and achieve was called neural plasticity. And so in order to achieve those kind of gains we want to be able to not fatigue. The muscle purposely just to build muscle mass now fatigue is our enemy and we want to be able to prolong the experiment as long as possible so how can we come up with new stimulation methods to maybe offset. The muscle fatigue that happens. Also you want to maximize the person's contribution and there are a lot of rehabilitation robots that are out there exit Skelton's and others where the person is strapped to a robot the robot moves around the person can take a nap or watch T.V. and just have their muscles moving around and that's not very beneficial others there are rehabilitation cycles where there's a motor that keeps the the desired cadence all the time and we have people who work out on those cycles and then try and do experiments in our lab they'll come and tell well we maximize our T three hundred I mean we I do train on that all the time but unfortunately they're being duped by the fact that the motor is doing all the work they're not really doing the work and they come in our lab and we're really making their muscles do the work all the sudden there's a sense of disappointment because they understand that before it wasn't their muscles that was doing the work it was really the machine that was doing all the work so we're really interested in questions like How can we maximize the person's active participation in the therapy and then you know we've now gotten to a point where we've done experiments on a wide range of people with a wide range of neurological conditions and some people are hypersensitive to the stimulation I mean you get near an electric field and they say this is pins and needles and I can't stand it and then there's just no way you're going to produce enough torque to do anything then you get to other people who they just say you know give me as much as possible it doesn't hurt in fact it's kind of a. Kind of sensation you know and and I'm doing these things that's cool Give me give me you know increases stimulation some people. Have Different their conditions cause their muscles to behave in a different way so we're kind of thinking of things now different torte production capabilities for example some people have a lot of fat content compared to their muscle to their their muscle fibers some people have have been long past their their injury and they just don't have the capabilities that they do compared to someone who's just recent past an injury for example and so we're now looking out talk to this about I'll talk about this at the end of the presentation on how can we accommodate for these changes in a person and how can we set it up to allow the person to do whatever they can do which is kind of coupled to the second question so maximizing what they can do but then. Building in allowances for what they can't do. So. Since we're controls and robotics people have the slide which is basically two sets of dynamics that are involved here one is the body segment on dynamics this is simple mass times acceleration kind of stuff or arranged dynamics and then you also have actually dynamics this is the electro chemical process that some bald there's only like phenomenological models for how the application of electric field causes chemical changes causes the production of force within the muscle fibers. So some of the phenomena that happen in there there are delays that are involved there nonlinear recruitment curves for how the muscle will start to contract there's fatigue so you can think of that as as a state dependent or time dependent controlled influence matrix if you will whereas over time you have less ability as the muscle fatigues the same input is less output. There are calcium dynamics that kind of are involved. In this whole process with fatigue and the generation of force and so we we have both sets of dynamics that are in play here when we're when we're doing muscle stimulation one of those. Things that talked about was the way and when you apply an electric field and you're expecting you know I apply my input I want to output immediately but there is this complex electrochemical process that happens in their body that causes the torque production and that yields this delay and the delay is on the order of like fifty to eighty milliseconds which may or may not sound a lot to you depending on your application field but I'll show you that it does make a difference in people. And so this is an input delay and input the Lays are an area that as a non-linear systems community there's still a lot of open questions and the problem is if you think about strategically if you know there's going to be a delay in the system well what can you do about it well as an engineer you would say OK well I'm just going to. Give the input now that I want to have happen T. seconds later where T. is the lay in the process so that then after the delay the control I want to apply happens at that time which is exactly the strategy everyone uses However if you have a linear system and you can predict how long it delays going to be because it's a linear system great then you can do that but what happens if it's a non-linear system and what happens if it's an uncertain on the your system it becomes much more complicated to predict when that delay is going to happen and in fact we found out through some experiments studies in our lab that as your muscle fatigues that delay gets larger Right so not only is it a uncertain delay in a non-linear system but it also changes in time and so. That becomes a very complicated for get muscles in the application domain that's a very complicated problem just from the calls at the vexed he kind of. Examination. So we've developed some controllers this is basically a component here and then this term is an interview all over the past and puts And one of the I guess interesting things that we did is that we integrate over a window where we don't know that. You would want to integrate over the window of the length of the delay in the system but we don't know the delay in the system so we're just approximating that delay with some estimate. And we were able to show. That we have exponential convergence to a residual error whenever we use these kind of controllers. Here's the lead up on our function for those controllers so you can see that we have some standard sum of squares kind of terms in there but then we have some other These are typically referred to as a functionals but there's a lot of innovation that comes in the design of these OK functionals in order to compensate or to enable us to design the controllers that can compensate for that delay. Here our experiment results where we implemented so the controller is a P.D. controller plus a still a compensation term and if we implement the P.D. controller you see a lot of bumpy behavior it's not it's not very smooth and the errors are you know they're getting to like fifteen degrees of error whereas if we implement the exact same controller on the same person with the same control gains everything's held the same except for now we turn on the delayed compensator term now you see a much smoother behavior by the controller the errors are much less the him put to the muscle is much smoother and. What it means in practice so this is an example of US applying the P.D. controller without the delay compensator and you see this jerky horrible this is an unstable behavior really it's not comfortable either by the way because you're going to shocked shocked shocked in a very sporadic way. Whereas same person saying gains same everything we just turn on the delay compensation term for this simple you know fifty to eighty millisecond delay and now you look nice smooth contractions which are also much more comfortable they also fatigue the muscle much less and we get much better behavior as we stall in those plots earlier so it really does. Make a difference in the application. We also are are looking at ways to adaptively compensate for the delay in the previous work we had a robust strategy to compensate for that on certain input delay here we're using a neural network we turn it into a partial differential equation where the delay is a spatial variable is transformed into a spatial variable and we do control on that. You also see kind of a sum of squares terms some neural network adaptation terms and then here's our functional which again there was a lot of innovation that came into the design of a functional the compensate for the delay. And then we've implemented that and we also get some nice results where we're actively compensating for the uncertain delay. And we've been working on that area for quite some time but I want to move on to some some other characteristics here is a is a reasonably good plot that describes muscle fatigue so in this I'm giving the same inputs throughout this three hundred second window but the axis on the left. It is torque output and you can see you know initially we were getting twenty five meters of torque but by the end we were getting like fifteen meters of torque So this is the time varying control influence if you will that's due to the muscle fatigue and we want to stimulate the muscle at a high enough frequency to get a smooth contraction. But the higher we stimulate the muscle the more rapidly we fatigue the muscle. And so what are some strategies that we can employ to get good results and effectively get high stimulation frequencies but not fatigue the muscle as quickly. And one idea is this idea of asynchronous stimulation so in a conventional single channel stimulation you know we would descend a pulse train to the muscle. But the idea between asynchronous stimulation is to spatially divide that channel up across the leg so here is someone who has a spinal cord injury and they have a single channel electrode So just one stimulation Pad whereas here we have multiple channels set up and we break this effective amount of energy up across different muscle fibers and there's going to be some crossover Yes but we at least try to spatially distribute the energy that we're putting in and the idea there is that you reduce the duty cycle that individual muscle fibers would receive but you're still these effectively of the same amount of energy is just distributed across multiple inputs and these are some experiments done on individuals with spinal cord injury these are error bars so you can see the results are pretty tight pretty pretty consistent in fact and the black curve is one where we stimulated a high frequency with the standard thing. Channel and that's comparable to the red curve which is the same effective stimulation frequency but distributed asynchronously and then the blue is the asynchronous curve that corresponds to the green one and so you can see the difference between almost exponential fatigue versus almost linear fatigue so that's a big deal for someone because and is statistically significant across a number of subjects that's what the slide shows the time for fatigue when you compare conventional like sixteen seconds mean in fifty seconds meaning to over one hundred seconds I mean that's a big deal in rehabilitation because if I can extend the number of repetitions that I do from fifteen seconds worth of repetitions to one hundred seconds repetition worth of repetitions the that's a lot more repetitions for my therapy. Now in order to do that asynchronous stimulation I distributed energy across the number of different channels and so that means that I am. I am switching between different actuators effectively is what's going on and we're discretely switching across the actual waiters and so my input is a discrete input switches from the different channels and then I have an input matrix that's also discretely changing between the different muscle fibers. In fact let me go ahead and go to this slide where. These are the torque out the torque output per input. Modulation signal for the different channels and you can see that not all channels across or not all spatial regions across the quadriceps in this case the same torque output and so you have to compensate for that as well. So going back we developed a P.I. the kind of air system. This is the. Open the bear system for that so you have the non-linear all arranged kind of dynamics in here a lot of the actuator dynamics are particular to the muscle or are captured in some a good term here. Some terms can be upper bound by constants others by. Higher order or quadratic functions of the state. And so for this controller we disagree is a simple sliding mode controller where we have a discontinuity there and then this is our resulting closed loop dynamics. And we were able to show exponential convergence. When switching across these different actuators Now one of the reasons why and we use the commonly a one off function to do that and one of the reasons why we're able to do this is we were always switching between a stable subsystem we were just each actuator was switching but each actuator was able to stabilize the system and so that made the problem a bit more simple. In the stability analysis though because we have discussed newbies on the right hand side you know we had a set value derivative and we went through all the mathematics in the details of that in order to get our exponential decay. And this is an implementation of the switching across the different channels of the. You see it's a nice smooth contraction even though there is the switching strategy across the channels to reduce the fatigue but we get very nice results when we when we stimulate the muscle in this way. In fact the same controller we implemented with a conventional single channel on a person same control gains everything and then we implement it just by dividing the channels asynchronously and here you can see the R.M.S. air goes down. But then it eventually creeps up because and here you can see the air are growing and the reason for the error growth is that we're just not able to have enough energy so the red curve which is the actual leg extension the height of the leg extension doesn't meet our desired blue you know leg extension trajectory Whereas when we stimulate asynchronously you see the error goes down and it doesn't increase because the muscle the fatigue over this period of time and our tracking errors are remaining constant and we're able to keep the same level of range of motion because we're not fatiguing the muscle. So again and is statistically significant across a number of subjects as what that shows So again by using switch systems theory and using those ideas to develop enough analysis of the controller to switch across the muscles we had a real physical impact in terms of the in terms of the therapy so here we also divide the muscle into a number of channels but for a different reason and this reason reason again you can see the muscle geometry changes significantly through the range of motion and if we had static pads which I mean there static pads. Then we would really see the effect of that changing muscle geometry and so to accommodate for that we are switching across the different stimulation paths to different channels state dependently as the function of the angle of the elbow in order to be able to maximize the torque production throughout the range of motion and I have another plot here in a moment but you can maybe see the different colors along this plot the different colors are when we need to switch to. Different stimulation channels in order to maximize the contraction the force by the bicep. And you can see you see kind of like some some wiggling there and it doesn't quite look like a natural contraction and that's caused by the electrical stimulation so we wrote one paper where we had three electrodes and we were switching discretely across the different electrodes to try and maximize the force and so these were plots of the force output as a function of the electrode position and so over here the blue electrode is going to give us more force production whereas over here the green electrode configuration is going to give us more force production so we switch across them and then we wrote a different paper where we blended the inputs in the time that a channel could give us some torque that was above a threshold then we blended that in with all the channels that could yield toward production and order to maximize the torque throughout the range of motion. But now I want to move on to cycling which is predominantly where our work is moving now. This person is not cycling on her own free will I mean her free will is just that she's going to sit there and allow us to stimulate her legs and let her be the be controlled by a computer but she's just passively sitting there in the computer is coordinating the stimulation across the different muscle groups in order to coordinate her limbs in order to have this nice cycling trajectory if I didn't tell you that it was a lie she's just peddling a bicycle and now imagine a child who had a spinal cord injury and is not able to ever ride a bicycle and then you give them this kind of technology and all the sudden they're able to ride a mobile bicycle it is a huge It's what I love about this field is that it combines both deep theory. Challenges as well as significant outcome and sense of accomplishment and impact in people's lives. In order to yield this kind of cycling we also have to switch across different muscle groups. Because in some parts of the cycle we would stimulate with our left leg versus our right leg. And when we're stimulating with a leg sometimes we will use our quadriceps more sometimes the user hamstrings more sometimes will use our glutes more and so there are different regions of the can the magic date when we would want to switch to different. Muscle groups and there's also a part of the gate basically when when the feet in the pedals are like this where I can stimulate the muscle really hard trying to generate torque but there's just not a lot of torque to generate when I'm in this configuration versus say in this configuration work and really apply a lot of torque right and so we don't want to stimulate the muscle when it's in these non. Favorable torque producing regions and so our and which of these grays shaded areas here so our options there are just not stimulate at all and stimulate enough when we can the kind of get through those regions and we do that and I'll show you some of that and then another way to do that is when I'm in these regions that it's not fruitful to cause the muscles to yield the sufficient amount to work well let's just put a motor in the loop and let the motor turn but not all the time to maintain cadence let's just intermittently use the motor in these regions when it's not efficient to use the muscle. So it's a closed loop kinematic chain you know what angle you know them all. The dynamics of the system you have torques produced by the cycle which are just standard mass times acceleration kind of things you have to. Work associated with a rider which is a passive torque which are kind of all arranged dynamics you have an active torque which is due to the electrical stimulation is also due to the person what they can voluntarily contribute if anything and if they can Boy we sure want them to because that's part of the therapy and then maybe there's some disturbances in each of those dynamics. And so the active torque This is where we can stimulate the muscles and we can yield cycling by by inducing torques through the muscles. Here you basically see that we have a switch system again it'll be state dependent switching across different muscle groups the right and left glute squads and hamstrings and then here is a state dependent again control influence matrix if you will. Where we have this torque transfer ratio this T.M. and basically if I were in a region where that tort transfer is above some threshold then we say that's a region where it's OK to stimulate that muscle otherwise we should stimulate that muscle because it's not going to produce useful torque. So we set up a switching scenario between switching between different muscle groups within the. Within the person and that's where we have our electrical stimulation input and then if we have a motor in the loop then we also have a switching condition for when to cut on the motor or not. In this one of our first experiments with cycling our control objective was just to maintain a desired cadence if you set your control objective up that way then you can write your open loop dynamics in this way with this term I just has a buy. Of. Linear non-linear terms in it but they can all be upper bound by a linear bound and that turns out to be very helpful mathematically. We have again controller and I'll come back to why all these controllers aside the controllers towards the end of the talk. And we're able to show. That when we're in the controlled regions with more stimulating the muscle where there's no motor in the loop here we're able to get exponential convergence that's no surprise we're using a slide in the controller. And then when we're in the uncontrolled regions where we just were just going to not stimulate the muscle we actually can have exponential growth and that's what you see here where this Lambie use a positive constant and this exponential growth is actually due to the fact that we can upper bound those collection of terms by a linear beyond. And here's kind of a plot of the strategy so when we are in regions where we're stimulating the muscle. Up in our function is convergent and when we're in regions where we're not stimulating the muscle then we can have exponential rise but as long as we're switching sufficiently fast. Then overall the magnitude of our all the up and all function is detaining and over all we can decay to ultimate ball or residual error and that's just going to be the be the case always because if you're switching to regions where you're not controlling it all you're always going to have some amount of growth and so you're always going to have this residual amount of error every cycle. And because we're switching between a stable controlled system and an unstable uncontrolled system we have to develop a time and that's what this condition. Here is which just says that based on how fast we converge when we're in the controlled region and how fast we blow up when we're in the uncontrolled region. And given ratios of the amount of time that were in those regions you can develop this overall region which says you have to converge at least this fast given knowledge of how long you're going to spin each of those regions OK So this is how a typical to all time condition. So we did experiments on a number of different people this is a healthy normal person and again it looks like they're peddling by themselves but but the machine is entirely paddling for them in fact sometimes we get people in with disabilities or neurological conditions and they're just so eager to help and to like calls this thing to work that they will they will try to control early on when we were naive in and told them just let the machine do everything I let let let us to me let your muscles in the appropriate way they would try to sneak in some contribution but we could instantly tell when they were doing that and in fact it would it would kind of disturb things which calls us to rethink the way we approach the problem which our show you and some of the later stuff so we had that just cadence tracking objective. We had this weird D.C. offset in our cadence is always behind a little bit and so and this picture just shows how we in one crank cycle how we distribute between switching between the different muscle groups and this is what it looks like with the whole thing normal person. Then when we stimulated someone with Parkinson's disease. One of the effects of their disease is that they were hyper sensitive to the stimulation so much so that we could not stimulate them enough to generate enough. Work for then be able to cycle and so with this person when we stimulate our healthy normals they don't know the desire to inject or they don't see any plots or anything we usually try to talk with them and distract them in fact but when this person could not just couldn't tolerate the amount of energy that we needed to put in to generate enough torque we showed him the trajectory and we said try to pedal and try to contribute so that you know you are producing some contributive torque so we don't have to stimulate you so hard and we did that he was able when he was contributing also there was no D.C. offset these gray boxes or when we also added in a lot of we changed gears on the bike basically on the fly and so it made it harder to paddle but you didn't really see too much of an effect on the error you see there's a little bit more energy that has to go and overcome the added resistance but here's someone with Parkinson's disease and you can tell that this side of his body was affected and you can see these curves look very some water to the person that's the healthy normal control but but these curves look very different from from the standard healthy normal. So when we compare you know what how good can you do right so we had people volitionally pedal you see that is our trajectory. You see your desired actual and you try to peddle the best you can to minimize the air and healthy normals they could get to you know less than two to R.P.M.'s of air with about in a desire to carry about fifty R.P.M.. And with electrical stimulation with them doing nothing we were a lie you know five plus or minus two rpm so he looks pretty good these are pretty reasonable results as far as it goes. But then when we look at the person with Parkinson's disease. It turns out they can also tell pretty good because you can cheat and you can one legged paddle right and so kind of one legged paddling or just doing the best he could possibly do. It wasn't as good as the healthy normal but it wasn't too bad but then when we took what he could do voluntarily and we added in the electrical stimulation on top of it I mean that's that's a pretty nice we can improve his ability to pedal when we're simultaneously stimulating him so that was kind of a nice result then we changed the control checked it to a position and velocity control objective now these terms in this bucket here now it contains some quadratic terms so we're not able to upper bound those linearly anymore we still use controller to control the muscles in the controlled regions we still get exponential stability but in the uncontrolled regions now because of those quadratic terms this quadratic nonlinearities now we have faster than exponential growth OK so this calls a lot of mathematics to have to deal with time conditions have things like natural logs of tangent functions. But really what it meant at the end of the day is the same scenario but because the escape is faster we just have to be more aggressive in our convergence when we're in the controlled regions and we can do well less time and uncontrolled regions. And again we converge to a ball and there's a lot of mathematics to dictate the size of the ball. And when we change the control objective it made the math a lot harder but then the results are better now we don't have these D.C. offset. Anymore in when we have our our people do the experiments so we still get you know within five R.P.M.'s of error but no D.C. offset then we did experiments where we added in the motor so so now our open loop dynamics contain both the electrical stimulation input and the motor and put. And then we have switching conditions for the electrical stimulation we have electrical stimuli we have a switch in conditions for the motor. And now we just always get exponential stability why because now there is no uncontrolled regions in the more because either we are controlling it through the muscle or we're controlling it through the machine either way we're just converging converging converging so in this way we were able to get overall exponential convergence and here's a picture of a strategy so now there's no uncontrolled region where we have the growth so we're just always decaying and here is a plot of the distribution of the control input between the different muscles in the different colors and in the black is when we needed to cut on the motor so you can see intermittently we're filling in for the muscles with the motor. Then we started to think about OK we want to do adaptive control and here was a big problem because all the previous controllers that you saw were also controllers because we had these twelve time conditions that we needed to me and we have to know how fast were converging in order to develop those dwell time conditions and so. You know unless you are doing some very exotic things to get parameter to in effect cation simultaneously it's hard to know how fast you're converging if you're using an adaptive controller because typically you have a knowledge strictly off and off function and you're only going to get asymptotic convergence but with a motor in the loop then you know there is no trial time condition because you're always switching between a stable subsystem and so with the motor in the loop and always switching between stable subsystems we said OK well now maybe this opens the door for us to be able to do our first adaptive control result and so the first Africa troll result that we tried was a repetitive learning controller So this is a type of a data control result that exploits the repetitive nature of the problem and with cycling it's very repetitive right I mean it's not exactly periodic. And so when we develop our controller. We segregate out some dynamics this W.D. Not coincidentally with more index and. We collect a bunch of terms in this this bucket which are functions of our desire trajectory which is very periodic And so now the strategy is to design our. Troll or this this new here to compensate for the for the periodic. Near terms that are uncertain the rest of the terms we crush with live in most control and robust feedback terms so here you see the adaptive feed for component proportional derivative feedforward term and then some sliding mode terms to compensate for the other uncertainties in disturbances in the system and we've developed this kind of a learning term which has a saturation function on it. In order to keep our repetitive estimate bounded. And then here now for the first time we have not an exponential convergence but now we have the limit as time goes to infinity we have this ass and product convergence of our cadence error system. The advantage that. Is that we're not having to dominate we're not having to use all of our controller to dominate the disturbances and uncertainties in the system with high gain high frequency feedback which over stimulates the muscle and fatigues it more rapidly we're able to offset some of the control energy to this feed forward term which is typically a lower frequency term which is typically a lower magnitude term it doesn't have these high gains multiplied by noisy sensor measurements and so has a lower frequency content in it and therefore the idea here is Stickley is that it would lead to lower muscle fatigue we haven't had enough time to test that out so right now is to remains a heuristic notion. And so here we have our sum of squares terms only up in a function and then here is our. Integral over the past period for our periodic learning terms. So we've implemented this controller on various people with different neurological conditions. We also implemented on our first person who is quadriplegic and so from my perspective it was a big step for us to move into a situation where we're doing experiments on someone who has no control of their body below the neck. And here is the repetitive learning controller on a healthy normal person again looks like cycling you can see this is the stimulator if you can see that little light coming on this little light of mine is when we're stimulating the muscle. And then this is a video that on one hand is like fantastic for me to see on the other hand I want to strangle the student who made it this is the person who's a quadriplegic you see them pet on the cycle this is an amazing triumphant video for us to record in the lab you can see is fingers they kind of just move based on rubbing against this other hand so he doesn't have any control over his body at all yet he's cycling but for some reason my student that is the hand in the video on the on the paddle and I just got this the other day and I said why did you put your hand on the pedal and I said well you know he has the sation and he said that is new is heard in a mobile bit and I said So why did you have your head on the pedal and I said I would like it made him feel better. Yeah but you killed my baby. So there's no reason for it in the have his hand on the paddle and I almost strangled him yesterday when I found out he did that but they he this guy's paddling completely through the computer making him part of the cycle which is which is just really nice to see. Kind of moving on and I know we're running out of time but some of our recent results. Kind of on this idea of let the person do as much as they can possibly do now we're setting up the switching zones where we have an aside. Distance region where we say here's the minimum speed that you should go in the cycle and in that region we're switching between electrical stimulation and maybe using a motor then we have this voluntary region where we say hey look if you can pedal beyond the pail and speed knock yourself out you should in fact we encourage you to pedal faster if you want to and then we reach an upper limit that says hey look some of these people who come in for these experiments again have a lot of enthusiasm and a lot of energy so much so that they could maybe be you know cause safety issues for themself or potential damage the equipment so they had a set of upper bound and on the upper bound now we have the motor fighting against them and basically increasing the resistance in the bike to a point where they're not going to fast anymore. But what's nice about it is that we have some just very preliminary experiments here where here's an assistive mode where the person is not pedaling fast enough and here you see lots of muscle stimulation going on in that region we also are steaming are switching back and forth on the on the motor. In some of the kinetic regions to also help and then the person paddles faster voluntarily then the desired objective now from a controls perspective it's hard to design this because you know a controller says we want to go to zero and if they're peddling faster than the desired speed that means ease not zero anymore and so the controller would want to fight against that but we've designed this in a way that it won't and we disallow them to stay in this uncontrolled region as long as they want to doing as much as they can and then we told this was a healthy normal person said OK now pedal pedal too fast and we set a threshold and now the motor cuts on and you can see the motor is providing resistance to fight against them so. I think this was either a C.D.C. paper or an A.C.C. paper. In the last cycles that we submitted really excited about this kind of war. We're also doing more work where we're looking at the human and the machine is coupled interconnected systems and we're taking a passive eighty based approach in the analysis and this is an admittance controller where the robot desire to inject lags behind. The person's trajectory by about ten degrees and we have the person is being electrically stimulated to fight against the machine and the machine is providing a desired and adaptive resistive force against the person but it's also admitting to the person so you develop this this passive ity base that mittens controller that allows the two the two systems to interact in a way to kind of optimize the work out here so here's the motor providing the resistance. And then here's the we're following the curves. As an explanation and then last slide. We're also moving into now upright walking kind of experiments where we have this step or treadmill we have instrumented a motorized so that we can do everything in close of control we also have a body weight support system and it's essentially a cycle that's upright but with having the person's body weight involved and having their arms involved that opens the door for new therapeutic outcomes and so we're kind of moving in this direction as well so it's a the three kind of directions that we're moving in are this kind of switching between these passive zones and allowing the person to decide to do as much as they can. Developing adaptive controllers. Even though we're switching among different subsystems. And these admittance. Kind of force control problems where we're looking at passive interactions or pass it using passively theory to look at the at mittens interactions between the person and the robot so that concludes my top. One hundred three. Questions. Where were using P. i D. plus something else. The only neural network controller that we implemented was on the time delay stuff and the role of the neural network was to learn what is the delay that the input delay which is uncertain. Yeah so that is you know kind of trial and error ad hoc tuning during the experiments so. You can increase the workload. And it is a real a robot exercise I mean if you're I got on the cycle Well if I go on a cycle maybe you're better and better say that I am if I own the cycle and they did experiments on me over the course of half an hour the next day I would feel as if I did a lot of a robotic exercise and I would have that kind of muscle soreness. So there is the potential to and in terms of like lifting stuff there's if you if you did the maximum amount that you could possibly lift and then you add electrical stimulation you can that's a lot more so it's interesting how much your mind limits what your body can do so. There. Are. I don't have any numbers but my guess would be horribly inefficient and that's not due to the fault of the muscle because the body does things incredibly efficiently but the way that the our window into using the muscle just externally and some people do internal stimulation they implant electrodes and they and their the group at Northwestern does this for example Case Western. But we don't want to get involved in the invasive experiments so we destabilize Turnley and when you stakes turn only and you're grabbing all the muscle fibers at once when you're stimulating you're really fatiguing the muscle rapidly and so this is probably the biggest problem to overcome in this area. Yeah I think that could be beneficial for sure you know we're right now we're just using. The joint angles which are easy to measure and then we developed a map that related say torque output based on the angle and. If we knew muscle length and muscle shortening velocity those kinds of things could improve our model and I'm a firm believer the more you know about the system you're trying to control the better job you can control it. OK. OK yeah that could be useful I mean we have read some literature about that at one time we also tried to use like a thermal camera to like blood flow in the muscle to maybe correlate that. I thought was a great idea the student really didn't completely buy into it and it kind of fizzled out but. Maybe they were like. The thing with you that the whole human with. The better part. So yeah I think that there's a cycle that goes on where we start with a theory that we know and we apply it to the application and then in in the in the practice of applying it we see things like I mentioned before where we were initially telling the people well hey don't tell just let us do let us actually it's your muscles and then we realized how dumb that was to do that because and we should other people tell as best they can and contribute as much as they can so then we went back to the drug drawing board and we said OK well how can we design a controllers now where that we allow the person to contribute as much as they can even if then contributing will cause our control errors to be non-zero can we develop the control structure to let that happen and so I think there's this perpetual cycle between doing the experiments and learning from the experiments thing going back and maturing theory and then and then repeat and repeat sound. Good. Is that a brass Tiger. Now maybe a bumble bee. Thank you.