WEBVTT 00:02.267 --> 00:04.307 Another really cool thing about Elliot 00:04.307 --> 00:06.187 that I want to highlight was his 00:06.967 --> 00:09.527 experience at the outside of Google. 00:11.267 --> 00:13.947 So he spent some time at Google 00:14.707 --> 00:17.947 founding their secret group of scientists 00:17.947 --> 00:19.967 that worked on eggs and skeletons, one of the 00:19.967 --> 00:22.087 co-founders of that for a couple of years. 00:22.367 --> 00:25.507 There was a company now called Skip that spun out 00:25.507 --> 00:28.327 of that effort. And he also spent a sabbatical at 00:28.327 --> 00:32.047 Boston by Anders A.I. Institute. So, man, when I look 00:32.047 --> 00:35.827 at Elliot's experiences, I'm kind of jealous a 00:35.827 --> 00:40.067 little bit, if I'm being honest, about the broad range 00:40.067 --> 00:42.227 of places that he's been that are all I've seen. 00:42.927 --> 00:43.787 Okay. 00:45.027 --> 00:47.547 Elliot's also one of my favorite tournaments. 00:48.447 --> 00:52.887 He's a very good designer and really cares 00:52.887 --> 00:54.747 about how he's going to feel. We're going 00:54.747 --> 00:57.107 to see a little bit about his work, I hope. 00:57.327 --> 00:58.547 Yes. Thanks. 00:58.807 --> 01:01.267 Thank you, Greg. thank you everybody for coming i'm 01:01.267 --> 01:04.167 really excited to be here um this is my second time 01:04.167 --> 01:06.827 visiting georgia tech so it's i haven't seen the 01:06.827 --> 01:08.727 campus i'd only seen the mechanical engineering 01:08.727 --> 01:11.187 building before but it's amazing it's a really great 01:11.187 --> 01:13.927 place is there any way to make that echo not happen 01:15.687 --> 01:19.187 what if i just turn the mic off and talk louder after 01:24.937 --> 01:27.277 you sure i can use that nice 01:33.347 --> 01:36.087 nice 01:42.627 --> 01:46.407 work all right now we can actually get started 01:46.407 --> 01:48.307 well i'm you know i'm very glad to be here. Thank 01:48.307 --> 01:50.777 you for coming. You guys are at a great place. 01:52.947 --> 01:54.787 So I want to tell you a little bit about 01:54.787 --> 01:56.527 kind of why I'm studying the things 01:56.527 --> 01:58.047 that I'm studying in my research group. 01:58.207 --> 02:01.007 And maybe I'll start off with some cool videos 02:01.007 --> 02:02.947 that you guys are, I'm sure, all well aware 02:02.947 --> 02:05.987 of. But let's think about bipedal robotics 02:06.107 --> 02:08.367 and what's been happening lately with humanoids. 02:08.367 --> 02:10.887 We have seen massive innovation 02:10.887 --> 02:12.627 that's really exciting, 02:12.907 --> 02:14.847 potential impact or even impact 02:14.847 --> 02:16.927 across diverse range of applications. 02:18.747 --> 02:21.847 and we see something kind of unique about the 02:21.847 --> 02:25.027 use of these robots, there's no human involved. 02:25.567 --> 02:29.567 Now, if we look at other areas of robotics, let's 02:29.567 --> 02:33.287 say wearable robotics, it's not quite the same story. 02:33.367 --> 02:36.587 It's a much different or, I would say, 02:36.587 --> 02:38.687 maybe much harder problem, and I want 02:38.687 --> 02:40.947 to blame that difficulty on the user. 02:41.207 --> 02:44.427 The difference between these two problem sets is that 02:44.427 --> 02:47.947 in wearable robotics, the robot is coupled to the 02:47.947 --> 02:50.467 wearer, and we have to interact with them and handle 02:50.467 --> 02:54.127 that in a way that's useful for them and promotes 02:54.127 --> 02:57.427 healthy biomechanics or mobility in the community. 02:57.427 --> 03:00.867 So we've seen some translation in wearable 03:00.867 --> 03:03.147 robotics, but something is different. And 03:03.147 --> 03:06.127 so I want to pin this on human interaction. 03:08.787 --> 03:12.887 so in my view wearable robotics have to solve this human 03:12.887 --> 03:15.367 interaction problem we have to be able to decode what 03:15.367 --> 03:17.967 the human is doing and interact alongside with them 03:17.967 --> 03:21.447 in a way that's helpful and in addition in our world 03:21.447 --> 03:24.987 in exoskeletons or robotic prostheses the way we assess 03:24.987 --> 03:29.427 these systems feeds right into how we build them and 03:29.427 --> 03:32.087 the success of these systems is also an open question. 03:33.547 --> 03:35.887 So if we want to build a successful 03:35.887 --> 03:38.247 wearable robot, it needs to be compact, 03:38.267 --> 03:40.707 lightweight, high performance. So lightweight 03:40.707 --> 03:44.267 being really, really important. It needs to respond 03:44.267 --> 03:46.507 to disturbances in the environment naturally. 03:46.687 --> 03:49.287 It needs to cooperate with their user, 03:49.287 --> 03:51.627 the person who's wearing it, with essentially 03:51.627 --> 03:53.687 little to no information transfer. 03:54.807 --> 03:56.627 Ultimately, these systems have to 03:56.627 --> 03:58.897 provide a benefit that outweighs their 03:58.897 --> 04:00.827 cost. I guess I'm not using the mic. 04:01.867 --> 04:03.687 And so in this talk, 04:11.457 --> 04:15.867 robotic systems, my research group, the 04:15.867 --> 04:17.257 Neurobionics Lab, is sort of structured with 04:17.257 --> 04:19.657 these two kind of high-level problems. 04:19.777 --> 04:21.617 This is actually really similar, I think, to 04:21.617 --> 04:23.757 kind of Greg's take. But the first and 04:23.757 --> 04:26.377 foremost, we study basic biomechanical science. 04:26.377 --> 04:28.457 And what I mean by that is we study the 04:28.457 --> 04:31.177 perception of people using these technologies. 04:31.177 --> 04:34.257 We study their biomechanics, their mechanical impedance, 04:34.257 --> 04:36.877 so their system dynamic properties that underlie 04:36.877 --> 04:40.257 gait, but we also build wearable robotic hardware. 04:40.257 --> 04:41.677 So we're going to talk about a few of these things 04:41.677 --> 04:45.177 today. But what I think is really important is the 04:45.177 --> 04:47.387 science that we're doing helps us build new 04:47.387 --> 04:49.857 technologies, and then we can take those new technologies 04:49.857 --> 04:52.157 and help them answer fundamental science questions. 04:52.157 --> 04:55.377 So you hopefully will get kind of a feel for that. 04:57.177 --> 04:59.877 So if we think about how are these technologies 04:59.877 --> 05:03.237 designed today, if you want to design a robotic prosthesis 05:03.237 --> 05:06.917 or an exoskeleton, where might you start? And there's 05:06.917 --> 05:09.897 kind of two philosophies today. The first is, 05:09.897 --> 05:12.937 we'll say, is expert-tuned. So it means a researcher is 05:12.937 --> 05:15.277 looking, sitting with the patient or the subject, 05:15.277 --> 05:18.817 looking at their gait and adjusting the behavior of 05:18.817 --> 05:22.517 that machine such that we see gait that we like. So 05:22.517 --> 05:27.037 that's almost always meant to replicate the kinetics and 05:27.037 --> 05:30.277 kinematics of healthy biological gait. And so this 05:30.277 --> 05:34.137 is the philosophy for a huge fraction of our field. 05:35.137 --> 05:38.057 Another kind of more recent approach is to do 05:38.057 --> 05:40.657 human loop optimization, where you sense a 05:40.657 --> 05:43.877 physiological variable, most often metabolic rate, and 05:43.877 --> 05:46.777 control the behavior of the robot to minimize 05:46.777 --> 05:50.117 that physiological variable, potentially minimize 05:50.117 --> 05:52.957 their metabolic cost. So the objective here, 05:53.417 --> 05:55.297 reduced metabolic rate. 05:55.417 --> 05:58.017 And that makes sense. If the device is working 05:58.017 --> 06:00.897 alongside you, we should see a reduced metabolic rate. 06:02.657 --> 06:05.397 But if you think about use of 06:05.397 --> 06:07.317 technology in the community, 06:08.107 --> 06:09.897 do we think we're always 06:09.897 --> 06:11.597 optimizing for one objective? 06:12.497 --> 06:15.677 Probably not. That's probably not how it works. 06:15.677 --> 06:18.697 We probably optimize across a host of objectives 06:18.697 --> 06:21.207 that come and go in terms of their priority. 06:21.337 --> 06:24.307 Metabolic costs certainly being important, 06:24.307 --> 06:27.037 but not the only dimension that matters. 06:27.337 --> 06:30.327 For example, you might care about speed, 06:30.327 --> 06:32.937 comfort, pain, stability, effort, other 06:32.937 --> 06:35.857 aspects of motion that would not be necessarily 06:35.857 --> 06:38.077 encapsulated by your metabolic rate. 06:38.747 --> 06:41.497 So our take has been to use user 06:41.497 --> 06:43.737 preference to get at this information, 06:44.127 --> 06:47.357 kind of let the user holistically be the optimization 06:47.357 --> 06:50.797 that subjectively selects through this abstract 06:50.797 --> 06:54.817 space of their experience. And we want to let the human 06:54.817 --> 06:57.877 brain do that and give that ability to the wearer. 06:58.617 --> 07:01.237 So we've been doing some work kind of 07:01.237 --> 07:03.457 on preference as a meta-criterion, 07:03.457 --> 07:05.777 sort of objective kind of higher up. 07:06.517 --> 07:11.137 And we've done some specific studies, some of 07:11.137 --> 07:12.617 which I'll talk about and some of which are just 07:12.617 --> 07:14.817 available in the literature if you want to see. 07:14.817 --> 07:18.117 But some things that we didn't know, for example, 07:18.257 --> 07:21.257 is what ankle-foot stiffness do people 07:21.257 --> 07:23.217 actually want from their prosthesis? 07:23.217 --> 07:25.697 So today's prostheses are just typically 07:25.697 --> 07:28.577 curved carbon fiber, foot-shaped springs. They 07:28.577 --> 07:31.017 have one set of mechanics that can't change. 07:31.017 --> 07:33.117 It's governed by the shape of that foot. 07:33.417 --> 07:36.037 So what stiffness do people actually want? 07:36.057 --> 07:39.037 In order to measure that, you might 07:39.037 --> 07:40.527 need to be able to vary stiffness. 07:41.257 --> 07:44.977 What drives user preference for people who have 07:44.977 --> 07:47.217 preferred stiffness in their prosthesis? What are 07:47.217 --> 07:49.537 they feeling that's giving them that sensation? 07:50.057 --> 07:52.977 How does user preference extend to ankle 07:52.977 --> 07:56.017 exoskeleton control? Can we wrap this concept in 07:56.017 --> 07:57.917 the loop? That project I won't talk about. 07:58.737 --> 08:00.897 And then how does clinician preference vary 08:00.897 --> 08:02.657 from user preference, which is especially 08:02.657 --> 08:05.877 important in the biomedical side of this world? 08:07.017 --> 08:09.897 Okay, so before we could address some of these questions, 08:09.897 --> 08:12.117 we had to build some new hardware. That's what I'm 08:12.117 --> 08:15.047 going to talk about first. So this is a variable 08:15.047 --> 08:19.097 stiffness prosthetic ankle foot that we developed some 08:19.097 --> 08:22.737 time ago. It has kind of like two important attributes. 08:23.057 --> 08:26.737 The first is it has a cam -based transmission. So 08:26.737 --> 08:29.717 this is a cam profile, a shaped piece of metal 08:29.717 --> 08:33.697 that rolls on the follower, and the follower then 08:33.697 --> 08:36.237 is deflected by that shaped piece of metal. And 08:36.237 --> 08:38.297 then we have a spring, a standard, in this case, 08:38.297 --> 08:41.517 titanium spring that we can, if you can imagine, 08:41.677 --> 08:44.307 this has pretty linear force displacement 08:44.307 --> 08:47.457 characteristics. But if we shape this cam profile 08:47.457 --> 08:50.607 correctly, we can arbitrarily decouple those 08:50.607 --> 08:53.957 mechanics. So this is sort of how it works. 08:53.957 --> 08:56.137 When the ankle joint is rotated, it deflects that 08:56.137 --> 08:58.517 spring. That spring is like a diving board. 08:58.977 --> 09:01.997 If we want to change its stiffness, we can adjust 09:01.997 --> 09:04.397 the support condition, which I'll show in second. 09:04.707 --> 09:07.197 So the point that I'm trying to make before is 09:07.197 --> 09:10.537 that this curved shape governs that torque angle 09:10.537 --> 09:13.177 deflection of that spring, that torsional spring. And 09:13.177 --> 09:16.057 we can make it, by shaping that metal correctly, 09:16.057 --> 09:18.797 have any set of passive torque angle mechanics. 09:19.437 --> 09:23.977 So then we're able to make this foot more or less 09:23.977 --> 09:26.437 stiff overall by adjusting the support condition of 09:26.437 --> 09:30.337 that cantilever spring. So very simple mechanism. 09:30.557 --> 09:32.937 It does have some design challenges 09:32.937 --> 09:36.037 associated with with multiple flat planes 09:36.037 --> 09:39.197 that have to be spaced and toleranced. 09:39.237 --> 09:41.897 But this is a very simple system that works really 09:41.897 --> 09:43.697 well. So this gave us the ability to vary 09:43.697 --> 09:45.997 stiffness from step to step, which had never been 09:45.997 --> 09:49.077 possible before in lower limb passive prostheses. 09:49.837 --> 09:52.837 It's quasi-passive, which means that it uses 09:52.837 --> 09:55.937 a small amount of electrical energy to sense 09:55.937 --> 10:00.097 and move that support condition, but it doesn't 10:00.097 --> 10:02.897 provide any net positive mechanical work. So 10:02.897 --> 10:05.437 it's not a powered prosthesis, it's passive. 10:08.187 --> 10:09.917 So we're just going to kind of talk 10:09.917 --> 10:12.057 through, highlight a couple of studies here. 10:12.557 --> 10:15.457 So this gets back to like what ankle foot stiffness do 10:15.457 --> 10:17.697 people actually want? So we don't know that because 10:17.697 --> 10:20.137 we've never been able to adjust the stiffness of a foot 10:20.137 --> 10:22.557 between steps. So the only way that we could know 10:22.557 --> 10:25.697 that before would be stop the treadmill. The subject dons 10:25.697 --> 10:28.577 the prosthesis, they put on a new prosthesis, start 10:28.577 --> 10:30.567 the treadmill back up, and then get their impression. 10:30.657 --> 10:33.837 So that time delay or that experience sort of loses 10:33.837 --> 10:36.217 their ability to provide high-quality information. 10:36.977 --> 10:40.637 So we took eight subjects, and we had 10:40.637 --> 10:43.097 them do five different activities. 10:43.217 --> 10:45.497 Stairs, incline, decline, level 10:45.497 --> 10:47.057 ground, so stair, ascent, descent. 10:47.377 --> 10:50.237 And we had them adjust the stiffness of their 10:50.237 --> 10:53.177 foot. So this is the torque angle relationship of 10:53.177 --> 10:56.337 that prosthesis, so it's a pure torsion spring. 10:56.647 --> 11:00.297 and we give them a dial and so they can adjust this 11:00.297 --> 11:03.477 dial that they're wearing and it will adjust the stiffness 11:03.477 --> 11:07.277 of the foot so between steps we're rotating through 11:07.277 --> 11:11.097 those torque angle curves so in every step they're 11:11.097 --> 11:13.737 experiencing one of those torque angle curves but it's 11:13.737 --> 11:17.037 continuous through that range so they have control with 11:17.037 --> 11:19.857 a dial we ask them to find their preferred stiffness and 11:19.857 --> 11:23.597 then we randomly reseed the location of the dial stop 11:23.597 --> 11:25.937 the treadmill and start it and ask them to do it again, 11:25.937 --> 11:28.217 and we can get a feel for how consistent they are. 11:29.597 --> 11:32.837 Here, we're looking at people's preferred stiffness, 11:32.957 --> 11:35.417 new meters per radian, or level ground, 11:35.417 --> 11:37.377 incline, decline, ascent, and descent. 11:37.837 --> 11:40.337 So this is for our eight subjects. This is their 11:40.337 --> 11:42.677 preferred stiffness, prosthetic foot stiffness across these 11:42.677 --> 11:46.337 activities. And here, we're normalizing by their 11:46.337 --> 11:50.317 level ground preference. So this says incline is 80% of 11:50.317 --> 11:54.237 level ground, decline is about 110%, ascent is about 11:54.237 --> 11:57.557 100% and descent is about 80 or 90. So it sort of tells 11:57.557 --> 11:59.917 us, that's telling us what fraction of their level 11:59.917 --> 12:01.917 ground stiffness do they prefer for these other 12:03.097 --> 12:06.237 activities. We did this with seven measurements. It tells 12:06.237 --> 12:10.357 us both their actual preference and their acuity to 12:10.357 --> 12:13.737 that preference by how the distribution of those data. 12:15.717 --> 12:20.657 The way prostheses are prescribed is by body weight 12:20.657 --> 12:24.637 and activity level. So those are the two dimensions that 12:24.637 --> 12:26.877 you tell them, and they will give you a prosthesis, 12:26.877 --> 12:29.557 a specific stiffness governed by your body weight and 12:29.557 --> 12:33.097 activity level. But what this is saying is people have 12:33.097 --> 12:35.997 very different preferences for what they want out of 12:35.997 --> 12:39.097 their prosthesis, yet we don't factor this in at all. 12:39.217 --> 12:43.897 We see a 30% change, maximum 30% change 12:43.897 --> 12:48.317 between decline and descent, which is two 12:48.317 --> 12:51.137 to three models of prosthesis difference. 12:51.917 --> 12:54.717 So that's two to three categories, which 12:54.717 --> 12:57.617 are these kind of metric meant to 12:57.617 --> 12:59.377 mean stiffness in the prosthetics world. 12:59.457 --> 13:04.497 So people want stiffnesses that are two to three 13:04.497 --> 13:06.997 categories different than what they're prescribed. So 13:06.997 --> 13:09.617 that's kind of a big deal. We're excited about that. 13:11.597 --> 13:15.327 also the within subjects variance maximum was 70 percent 13:15.327 --> 13:17.557 so that just says people don't all want the same 13:17.557 --> 13:20.517 thing and then even within a subject they might want 13:20.517 --> 13:22.897 very different things across these activities so it's 13:22.897 --> 13:25.517 important to be paying attention to the mechanics of a 13:25.517 --> 13:28.317 prosthesis which we haven't really been able to do 13:28.317 --> 13:31.377 because they've always been foot-shaped carbon springs 13:32.887 --> 13:35.177 okay we looked at what drives user preference 13:35.177 --> 13:37.957 so what what are the underlying biomechanical 13:37.957 --> 13:40.237 factors that might govern this sensation of 13:40.237 --> 13:42.537 this is my preferred stiffness. So to do this, we 13:42.537 --> 13:45.417 had seven people walk at fractions of their 13:45.417 --> 13:47.777 preferred stiffness at three different speeds. 13:48.157 --> 13:50.337 So we're looking at walking speeds, 13:50.337 --> 13:52.897 essentially self -selected, fast and slow. 13:53.877 --> 13:56.497 And then we're looking at, at those speeds, 13:56.497 --> 13:58.317 we're looking at different stiffness settings. 13:58.627 --> 14:00.397 What is their, you know, at their preferred 14:00.397 --> 14:02.737 stiffness, plus and minus 15 percent, plus 14:02.737 --> 14:05.057 and minus 30 percent. So what we're looking 14:05.057 --> 14:07.347 looking for is something that has an optimum 14:07.347 --> 14:09.637 around their preference, that's the game. 14:10.177 --> 14:13.677 So to do that, we measured metabolic rate and 14:13.677 --> 14:16.637 full biomechanics. So something like 90 variables. 14:17.457 --> 14:21.497 And we looked through these variables to see, do 14:21.497 --> 14:25.787 any of them have a quadratic profile where the minimum 14:25.787 --> 14:28.157 of that quadratic function or maximum is within 14:28.377 --> 14:30.737 their preference? So that's what we're looking for. 14:30.937 --> 14:33.237 We saw lots of potential 14:33.237 --> 14:34.637 trends, as you might imagine. 14:36.137 --> 14:39.437 Here we're looking at subjects' metabolic rate as a 14:39.437 --> 14:42.097 function of difference from their preferred stiffness. 14:42.297 --> 14:46.117 And the main takeaway from that is no change, which 14:46.117 --> 14:48.517 is against what the literature would say. If we see a 14:48.517 --> 14:52.317 massive change for speed, speed has a huge effect on 14:52.317 --> 14:55.797 metabolic rate, but changing the stiffness did not. 14:55.887 --> 14:58.737 So we don't think it's metabolic 14:58.737 --> 15:01.137 rate that people are keying into. 15:01.657 --> 15:04.817 We saw many trends that look like this. So 15:04.817 --> 15:07.497 this is just looking at affected ankle peak power. 15:07.497 --> 15:10.917 So peak power provided by the prosthesis. 15:11.257 --> 15:13.437 You might think if they're missing their ankle 15:13.437 --> 15:16.117 joint, they might want to maximize that power, 15:16.117 --> 15:18.517 maximize the energy provided by this passive system. 15:18.887 --> 15:20.837 But that's not what they do. If that was the 15:20.837 --> 15:22.837 case, they would go softer and softer and softer. 15:22.897 --> 15:24.917 So that's what that's saying. the softer 15:24.917 --> 15:27.737 you go, the more energy it stores, which 15:27.737 --> 15:30.017 makes sense. So they're not trying to 15:30.017 --> 15:32.777 store or return as much energy as possible. 15:33.217 --> 15:37.957 We only saw one variable that had this optimum where 15:38.257 --> 15:40.617 it's a second order function where the 15:41.527 --> 15:44.797 optimum is within their preference. And that 15:44.797 --> 15:47.917 was kinematic symmetry between their 15:47.917 --> 15:50.857 biological ankle and their prosthetic ankle. 15:51.057 --> 15:53.677 And that was the only variable that 15:53.677 --> 15:55.997 mapped to, that was optimized at their 15:55.997 --> 15:58.177 preference, which is sort of amazing. 15:58.917 --> 16:01.657 So we don't know exactly what that means 16:01.657 --> 16:03.677 yet or the significance of that. Certainly, 16:04.077 --> 16:07.297 symmetry is good. In clinical outcomes, the 16:07.297 --> 16:10.777 clinicians would be very, very positive on seeing more 16:10.777 --> 16:14.277 symmetric gait, but we don't know exactly what it 16:14.277 --> 16:17.177 is that's causing that to be the only variable. 16:17.537 --> 16:21.657 So we're still looking at kind of this area of research. 16:21.657 --> 16:23.617 We're looking at above knee amputations instead 16:23.617 --> 16:26.497 of below knee, but we're trying to kind of understand 16:26.497 --> 16:28.917 what is it that people are actually caring about. 16:29.757 --> 16:32.437 And it's not maybe the things we thought it was. 16:32.967 --> 16:37.077 So I'm going to kind of transition a little bit 16:37.077 --> 16:40.017 here. We developed a prosthetic ankle that can vary 16:40.017 --> 16:42.357 its stiffness from step to step. We did that with 16:42.357 --> 16:46.117 a really small mechanism and we used it as a tool 16:46.117 --> 16:49.457 to assess or investigate prosthetic mechanics. 16:50.597 --> 16:54.297 We've also developed use this information to study 16:54.297 --> 16:56.877 roles of clinician preference. I think that's a really 16:56.877 --> 17:00.037 important concept, that one of those references is 17:00.037 --> 17:03.677 that paper. And we've developed a version of this foot 17:03.917 --> 17:06.397 that can recycle energy. 17:06.847 --> 17:10.187 So I told you that it was a spring and a cam, and 17:10.357 --> 17:13.477 the cam deflects the spring, and we can create torque 17:13.477 --> 17:16.997 angle mechanics that we want. But maybe now imagine 17:17.287 --> 17:19.837 that when I rotate the ankle joint and the cam 17:19.837 --> 17:22.997 deflects the spring, that once it's maximally deflected, 17:22.997 --> 17:25.297 I switch the cam profile out to a different cam 17:25.297 --> 17:28.037 profile. And now I can return that energy in a different 17:28.037 --> 17:30.337 way or do interesting things with that energy. 17:30.337 --> 17:32.817 So we're going to talk about that for a minute. But 17:32.817 --> 17:34.317 we're going to talk about that in the context of an 17:34.317 --> 17:36.917 orthosis or exoskeleton version of this technology. 17:37.437 --> 17:38.957 This is our variable stiffness 17:38.957 --> 17:40.637 orthosis or exoskeleton. 17:40.957 --> 17:44.097 It's the same thing, it's just much smaller 17:44.097 --> 17:46.497 since it's only providing a small amount of 17:46.497 --> 17:48.717 the torque since the person has as the ankle 17:48.717 --> 17:50.897 joint, but it's essentially the same thing. 17:50.897 --> 17:54.137 You can see cam profile, spring, slider. 17:57.097 --> 17:58.677 It's about 900 grams. 17:58.677 --> 18:00.577 It can provide about 80 newton meters. 18:01.497 --> 18:04.237 And we developed this mechanism, 18:04.237 --> 18:06.697 which we called the DESR mechanism, 18:06.797 --> 18:08.697 decoupled energy storage and return. 18:08.837 --> 18:10.317 Just interesting. 18:10.917 --> 18:13.477 So the way we do this, so now we're kind of looking 18:13.477 --> 18:18.317 like right here at the kind of back of it, and we see this 18:18.317 --> 18:21.937 shape here, this long part, that's the cam profile. 18:22.377 --> 18:25.337 And then this center part is a part 18:25.337 --> 18:26.957 that switches out back and forth. 18:27.177 --> 18:30.497 And that center part has different sets of mechanics. 18:30.497 --> 18:34.237 And we can push these little cam pieces around with 18:34.237 --> 18:37.397 small permanent magnets, because when they transition, 18:37.397 --> 18:40.177 the follower is on another place of the cam profile and 18:40.177 --> 18:42.657 they're unloaded. So you'll see this a little more 18:42.657 --> 18:45.677 carefully. So this is two cam profiles that interchange. 18:45.837 --> 18:48.197 It looks something like that, for example. 18:48.597 --> 18:51.457 So heel contact would be here. 18:51.717 --> 18:55.417 You would go heel contact, foot flat, the cams 18:55.417 --> 18:58.577 would switch. Then it's riding up the orange cam 18:58.577 --> 19:01.197 up to here, and then this would be push off. And 19:01.197 --> 19:04.697 what happens is you recycle this with collision. 19:05.037 --> 19:08.177 That's negative energy. This is positive energy. 19:08.217 --> 19:12.097 So we can fake a powered system by capturing 19:12.487 --> 19:15.117 collision energy, storing it in the mechanism, 19:15.117 --> 19:17.797 and providing it later. And so that's what 19:17.797 --> 19:20.137 this sort of cam, interchanging of cams allows. 19:20.817 --> 19:23.157 Energy recycling, so that's pretty cool. 19:23.317 --> 19:25.657 But it also could allow changing of the 19:25.657 --> 19:27.957 equilibrium positions of these torsion springs. 19:28.137 --> 19:32.357 So many people who wear orthoses, the primary 19:32.357 --> 19:34.557 function of that orthosis is to keep them from 19:34.557 --> 19:36.997 stubbing their toes in swing phase. Their 19:36.997 --> 19:39.637 foot will drop and then hit the ground. That's 19:39.637 --> 19:41.517 a big problem. So an orthosis keeps that up. 19:42.357 --> 19:46.037 This type of orthosis could automatically dorsiflex 19:46.037 --> 19:49.377 the foot up in the air and swing and then switch back to 19:49.377 --> 19:52.197 a regular orthosis on heel contact. So there are some 19:52.197 --> 19:55.137 really interesting mechanical things we can do here. 19:55.237 --> 19:57.157 We can do extreme positive and 19:57.157 --> 19:59.597 negative stiffness, which is cool. 20:00.597 --> 20:03.797 That takes some explaining, but that we can 20:03.797 --> 20:07.557 render stiffnesses from this device higher 20:07.557 --> 20:10.357 than any constitutive stiffness in the system. 20:13.057 --> 20:16.937 So we currently have this work is 20:16.937 --> 20:19.007 right about to come out in JNER. 20:19.617 --> 20:23.137 And then we have two optimized CAM profiles for 20:23.137 --> 20:25.357 people with pathology that we're studying now. 20:28.047 --> 20:30.607 Okay, now we're going to switch gears to a machine 20:30.607 --> 20:34.957 design project. This is a fun exploratory project. 20:36.787 --> 20:38.137 I'll just keep going. I have a 20:38.137 --> 20:39.337 few more projects to talk about. 20:39.517 --> 20:42.977 Okay, this project is about exploring 20:42.977 --> 20:44.707 a new way to design an exoskeleton. 20:45.787 --> 20:47.857 So we want a machine that's going to provide 20:47.857 --> 20:51.177 positive mechanical power to the body that you wear. 20:51.817 --> 20:53.757 But we want to make exoskeletons 20:53.757 --> 20:56.057 that don't look like exoskeletons. 20:56.277 --> 20:59.757 So on the left, you see this chat GPT generated 20:59.757 --> 21:02.997 schematic. mechanic, this is showing two common styles of 21:03.337 --> 21:07.057 exoskeleton actuation. One is to put a motor right on 21:07.057 --> 21:09.677 the joint in the sagittal plane, and the other is to 21:09.677 --> 21:12.477 run cables up to a backpack and pull on these cables. 21:12.577 --> 21:17.137 Those are pretty much the two main styles these days. 21:17.437 --> 21:21.657 So these exoskeletons generally always look 21:21.657 --> 21:23.717 like exoskeletons. They have a pretty clear, 21:23.777 --> 21:26.677 distinctive look to them. And so this idea is, can 21:26.677 --> 21:29.597 we restructure the way we build the actuator 21:29.597 --> 21:31.897 such that it doesn't look like an exoskeleton? 21:31.967 --> 21:33.617 And here what we're going to do is try to 21:33.617 --> 21:36.537 make an actuator with an extreme hollow bore 21:36.817 --> 21:38.757 that we can put our limb through. 21:39.477 --> 21:42.307 But in order to make it so that it's not humongous, 21:42.797 --> 21:44.977 we want to make the motor open and close. 21:46.297 --> 21:49.097 So one idea is to mount the actuator mass around 21:49.097 --> 21:51.637 the leg so the actuator spins around the body. 21:52.117 --> 21:54.497 You can keep everything close, maybe. 21:55.017 --> 21:58.017 Mount it into a shoe or something. So 21:58.017 --> 22:00.657 hoop-style actuations around limbs. 22:01.037 --> 22:04.397 And if we can take a motor, so we built 22:04.397 --> 22:05.837 one of these that you could put your leg 22:05.837 --> 22:08.737 through. And the largest dimension is this, 22:09.187 --> 22:12.757 from a bottom heel to the top of your foot. And 22:12.757 --> 22:15.437 so that diameter is non -trivial. That means the 22:15.437 --> 22:17.577 motor is this big. And so we have this. So we have 22:17.577 --> 22:20.137 an exoskeleton that has a giant hoop around your 22:20.137 --> 22:23.217 leg that will provide power. But it was so large 22:23.217 --> 22:25.577 that there's no chance of that ever being viable. 22:25.727 --> 22:28.077 So the only way we could think to get it closer 22:28.077 --> 22:30.997 was to split it in half and have it open and 22:30.997 --> 22:32.917 close. So that's what I'm going to talk about. 22:33.877 --> 22:36.477 So we developed an extreme hollow bore brushless 22:36.477 --> 22:39.097 motor that can open into two halves. So 22:39.097 --> 22:40.857 that's a picture of it. That's a standard motor 22:40.857 --> 22:43.837 on the left and the final design on the right. 22:44.277 --> 22:46.217 So we took a commercial motor 22:46.217 --> 22:48.917 from T-Motor, the MN-1005. 22:48.937 --> 22:50.077 That's this. 22:50.567 --> 22:52.277 And we tried to see if we could make 22:52.277 --> 22:53.877 one of those that could totally split in 22:53.877 --> 22:55.497 half and come together and still work. 22:57.967 --> 23:00.237 So I'll talk for a minute about how we did this. 23:00.477 --> 23:02.957 So we took the stator core. So this is 23:02.957 --> 23:06.657 the inside of an exterior rotor motor. 23:06.807 --> 23:07.917 Copper would be round around 23:07.917 --> 23:10.417 this. We wire EDM'd it in half. 23:11.057 --> 23:12.857 So we cut this thing in half. 23:13.257 --> 23:16.057 and then we had to wind it like a standard 23:16.057 --> 23:19.597 brushless motor and we understood the winding pattern 23:19.757 --> 23:22.377 and we had to wind it except for it can't if 23:22.377 --> 23:24.477 it's going to open here you know the windings can't 23:24.477 --> 23:25.917 cross there so that was really what we're 23:25.917 --> 23:28.877 solving we have to stop the windings here bring it 23:28.877 --> 23:32.137 around they go here go back and then come this 23:32.137 --> 23:36.237 way so we rewound this motor such that it only is 23:36.997 --> 23:39.737 or it's not connected on one side. 23:40.177 --> 23:43.117 It's a delta-wound motor with 22-gauge 23:43.117 --> 23:45.737 copper winding, 18 coils per tooth. That's 23:45.737 --> 23:48.057 the winding pattern that's shown here. 23:50.217 --> 23:52.357 We tested the back EMF. 23:52.357 --> 23:56.157 So this is measure the voltage 23:56.157 --> 23:57.857 between two leads and spin the motor. 23:58.037 --> 24:00.957 So this is the split version. This is the unsplit 24:00.957 --> 24:04.237 version, and we see essentially identical back EMF 24:04.237 --> 24:07.797 waveforms. They maintained their sinusoidal pattern. 24:07.797 --> 24:09.037 All of that looks good. 24:11.017 --> 24:12.877 This is pretty much, I think, all I'm 24:12.877 --> 24:15.177 going to show in the design. But the 24:15.177 --> 24:18.017 way we did this is it's a structure, 24:18.777 --> 24:21.717 an internal structure that mounts to the stator core. 24:21.977 --> 24:23.957 And then it's a set of bearings, 24:23.957 --> 24:25.387 which we call the bearing of bearings, 24:25.577 --> 24:30.177 because a large diameter bearing can't open. So 24:30.177 --> 24:32.897 we had to solve that. So what we did was use many 24:32.897 --> 24:34.937 small radial bearings that run around a race 24:34.937 --> 24:37.897 and so that's these they're mounted to this red 24:37.897 --> 24:40.357 structure that mounts to the stator core and then 24:40.357 --> 24:44.777 this black structure holds the flux ring and all 24:44.777 --> 24:46.877 the magnets and it's what's actually rotating. 24:47.237 --> 24:50.357 It has a closure mechanism that's like a 24:50.357 --> 24:54.997 watch so it it looks exactly like this. It 24:54.997 --> 24:56.937 has a closure mechanism that does this. 24:58.237 --> 25:00.237 and locks and closes, so you can't 25:00.237 --> 25:02.437 really see it, but it's here. 25:04.997 --> 25:07.797 It's like a folding mechanism. 25:09.217 --> 25:11.657 So that's a motor that can split in half. 25:13.217 --> 25:15.357 We had to add an encoder to it. We had to add two 25:15.357 --> 25:18.437 encoders because we were not sure if it's going to lose 25:18.437 --> 25:21.077 counts as it goes around. We had to take the encoder 25:21.077 --> 25:23.717 and split it in half. So now we have to read those 25:23.717 --> 25:26.837 encoders counts and ensure that it's not losing counts 25:26.837 --> 25:30.557 across the split. So we use like a standard magnetic 25:30.557 --> 25:33.137 encoder and split it on the poles and then machined 25:33.137 --> 25:36.317 out the steel behind it so it's a little bit lighter. 25:36.577 --> 25:40.537 And so the motor is 304 grams without the encoder 25:40.537 --> 25:43.737 and 360 grams with the encoder. So that's the 25:43.737 --> 25:45.577 full setup. You can see the kind of packaging 25:45.577 --> 25:49.277 and read head, but that's just adding packaging 25:49.277 --> 25:51.417 the encoders. It turned out we don't need two 25:51.417 --> 25:54.217 encoders. It actually doesn't miss any counts, 25:54.587 --> 25:56.817 which I was surprised by. We split it right 25:56.817 --> 25:58.897 at the pole, and the encoders are smart 25:58.897 --> 26:01.237 enough to interpolate, I guess, across that. 26:03.037 --> 26:05.737 So now I want to tell you, like, oh, it works. So we 26:05.737 --> 26:08.037 tested it in a test bed here. This is kind of a 26:08.037 --> 26:11.437 standard dyno, motor dyno, two motors kind of running 26:11.437 --> 26:13.477 against each other through a contactless torque sensor. 26:14.537 --> 26:17.617 They share a large brushless battery. 26:18.917 --> 26:21.117 So we're just going to talk through a few of the 26:21.117 --> 26:24.077 properties. Here we're looking at torque constant. So 26:24.077 --> 26:26.657 torque as a function of current, that's Q axis current. 26:27.577 --> 26:31.117 It was the same. So the fact that we split the motor, 26:31.117 --> 26:33.917 rewound it differently, none of that mattered. We 26:33.917 --> 26:36.137 still have a torque constant that agrees within 2%. 26:36.897 --> 26:39.477 This we're looking at kind of stiction and viscous 26:39.477 --> 26:42.517 damping. So velocity and current, This is the 26:42.517 --> 26:46.037 stiction. That's the viscous damping. And so it has 26:46.037 --> 26:49.057 a lot more viscous loss. And that comes from the 26:49.057 --> 26:52.257 many bearings that are in it. So it runs, you can 26:52.257 --> 26:55.177 see it has a lot more viscous loss compared to the 26:55.177 --> 26:58.417 unsplit motor. That's one of the clear downsides. 26:59.557 --> 27:01.937 Now we're looking at efficiency maps. 27:02.057 --> 27:05.037 So this is current and voltage and the efficiency 27:05.037 --> 27:08.157 at that That regime for both the split motor and 27:08.157 --> 27:12.727 the unsplit motor scaled to 100% to minus 100%. 27:14.727 --> 27:17.017 Visually, you can tell they're very similar. 27:17.017 --> 27:18.677 So even though, so there's a lot more 27:18.677 --> 27:20.577 winding, that creates a lot more resistance. 27:21.117 --> 27:26.077 It is about 4% less efficient, but that extra 27:26.077 --> 27:28.377 copper didn't seem to add too much resistance. 27:31.677 --> 27:34.117 We do lots of thermal characterization. This is 27:34.117 --> 27:36.657 kind of the last step of our analysis pipeline. 27:36.837 --> 27:39.037 So we did our thermal characterization. 27:39.037 --> 27:41.037 These are GIFs, actually, which I think are 27:41.037 --> 27:42.957 really cool. You'll see it. Well, you can only 27:42.957 --> 27:44.557 see them for like a second when they start over, 27:44.557 --> 27:47.917 but this is how we both understand the winding 27:47.917 --> 27:51.177 type and quantify the thermal properties. 27:51.337 --> 27:55.037 So we have a thermal physics model that calculates 27:55.037 --> 27:57.737 or estimates the winding thermal resistance, 27:57.737 --> 27:59.737 housing thermal resistance, and capacitances. 28:01.757 --> 28:06.017 so we've calculated this now we have essentially all 28:06.017 --> 28:09.497 the information we need to design with these types 28:09.497 --> 28:12.517 of devices we also made a really fun rendering we 28:12.517 --> 28:14.577 actually made that we made this with chat gpt and in 28:14.577 --> 28:17.517 the sense that we asked chat gpt to to create us a 28:17.517 --> 28:20.877 apple style commercial for this technology and then 28:20.877 --> 28:22.997 we had to do the rendering we told it we had to tell 28:22.997 --> 28:25.637 us exactly the shots there's actually a voiceover 28:25.637 --> 28:29.077 that's my voice that's ai recorded it's an ai version 28:29.077 --> 28:32.677 that it created all the all the voicing for it but 28:32.677 --> 28:34.917 this was just sort of a fun thing that we made but 28:35.337 --> 28:37.077 this is just a cool rendering of 28:37.077 --> 28:38.697 the design that kind of shows it off 28:43.057 --> 28:44.957 so yeah there it is running you 28:50.127 --> 28:52.667 can see there's a little bit of like run out 28:52.667 --> 28:54.987 on the encoder but it doesn't seem to matter 28:58.967 --> 28:59.887 Okay. 29:00.217 --> 29:01.597 I have two more projects. 29:02.827 --> 29:04.107 I think we're good on time. 29:04.467 --> 29:07.807 Okay. So this is a project that's the open 29:07.807 --> 29:11.247 source robotic leg. This is the largest and longest 29:11.247 --> 29:13.747 running project in the group and also has 29:13.747 --> 29:16.517 had the most federal funding. So it's heavily 29:16.517 --> 29:18.487 supported from the National Science Foundation. 29:19.007 --> 29:22.627 So this was the idea that people are 29:22.627 --> 29:24.287 investigating control strategies, 29:24.287 --> 29:26.227 developing controllers for robotic prosthetic 29:26.227 --> 29:27.567 legs. That's one of the main reasons you 29:27.567 --> 29:29.727 don't see them in the public is because we don't 29:29.727 --> 29:31.187 know how to control them well enough yet. 29:31.307 --> 29:35.067 And to tackle that barrier, people were building their 29:35.067 --> 29:37.707 own robotic legs first and then studying the control. 29:37.707 --> 29:40.667 And you can probably imagine how difficult and 29:40.667 --> 29:43.967 expensive that would be. But in addition to that, it made 29:43.967 --> 29:46.447 it so it was very difficult to compare across 29:46.447 --> 29:50.367 researchers and to act as a field. It would be really hard 29:50.367 --> 29:53.307 to know if the results from this leg compared to that 29:53.307 --> 29:55.647 leg are due to the control or due to the hardware. 29:55.647 --> 29:58.387 And so that is a challenge if we're 29:58.387 --> 30:01.447 trying to work together to solve these 30:01.447 --> 30:03.327 challenges for the people who need them. 30:03.447 --> 30:07.427 So we developed an ecosystem of hardware 30:07.427 --> 30:10.567 and software tools to help people get started 30:10.567 --> 30:12.487 studying the control of robotic legs. 30:12.507 --> 30:15.027 So it's a design, the open 30:15.027 --> 30:17.087 -source leg V2, that's this. 30:17.647 --> 30:21.787 It's Onshape open-source CAD is on that QR code. 30:22.507 --> 30:26.887 The next one is the Raspberry Pi imaging tool. That is 30:26.887 --> 30:32.647 an amazing tool. It's a continuous integration tool 30:32.647 --> 30:35.087 that will build an operating system for you that you 30:35.087 --> 30:38.047 can then flash to a Raspberry Pi with tons of enhanced 30:38.047 --> 30:40.927 networking and communication features. It's amazing. 30:40.927 --> 30:44.367 We have the website, which is opensourceleg.org, and 30:44.367 --> 30:47.307 we're releasing embedded systems. That was the weakest 30:47.307 --> 30:51.267 link now is that we need shared embedded systems. 30:51.387 --> 30:53.967 So we use the Raspberry Pi compute module, and we 30:53.967 --> 30:57.607 have a set of interface boards that allow expansion 30:57.607 --> 31:00.027 and functionality with not only the open source leg, 31:00.027 --> 31:02.807 but sort of general sensors and actuators. We have 31:02.807 --> 31:05.607 a whole host of sensors and actuators available. 31:06.067 --> 31:10.087 So this is the full set of things, but 31:10.087 --> 31:11.627 now I can tell you a little more about the 31:11.627 --> 31:13.627 hardware and the details of the project. 31:14.247 --> 31:16.427 This is the open source leg. It's a modular two 31:16.427 --> 31:18.727 degree of freedom robotic leg. It can be used 31:18.727 --> 31:21.047 as a knee or an ankle together or separately. 31:23.367 --> 31:25.587 If we look inside, it has a kind 31:25.587 --> 31:27.137 of a few things that are relevant. 31:27.967 --> 31:30.507 The first is we use an actuator. 31:30.987 --> 31:34.127 The actuators that we use from CubeMars have a 9-to 31:34.127 --> 31:36.447 -1 in them already, so that was really nice. Those 31:36.447 --> 31:39.447 weren't available in the first generation we built, 31:39.447 --> 31:42.327 that was not a product yet. So we use the first 31:42.327 --> 31:45.527 stage from this 9-to-1 on the actuator, then we add 31:45.527 --> 31:48.287 a second stage, and that's what this is. So there's 31:48.287 --> 31:51.887 two identical second-stage belt drive transmissions 31:51.887 --> 31:55.627 that bring the ratio up to about 45 or 47 to 1. 31:57.687 --> 32:00.647 So that's a, I'll say, low -ish transmission ratio. 32:00.797 --> 32:02.687 It's lower than they have been used in the 32:02.687 --> 32:05.727 past, so it facilitates control in that sense. 32:05.967 --> 32:08.827 Belt drives are easy and cheap, quiet. 32:09.907 --> 32:12.187 We have onboard sensing and batteries 32:12.187 --> 32:14.087 that have BMS. I think that's important. 32:14.827 --> 32:17.927 And so fully integrated sensing includes 32:17.927 --> 32:20.007 a load cell, 6-degree freedom load cell, 32:20.187 --> 32:24.007 IMUs on the joints, encoders at the joint and at 32:24.007 --> 32:28.427 the motor, pretty much all motor state information, 32:28.627 --> 32:29.727 everything you would need to 32:29.727 --> 32:32.047 know to do control, maybe. 32:34.287 --> 32:37.107 So now we also, with the second generation system, 32:37.107 --> 32:39.067 the knee and the ankle share drivetrain parts. 32:39.067 --> 32:42.527 So that's really helpful, I think, for people. 32:43.687 --> 32:46.707 And they both include the concept of series 32:46.707 --> 32:49.347 elasticity, which I'm guessing you guys have heard 32:49.347 --> 32:52.147 about or know about, but I'm going to talk 32:52.147 --> 32:55.907 about it for a moment. And I'd say it's not a solved 32:55.907 --> 32:58.667 problem in our solved. We all don't have the 32:58.667 --> 33:01.147 same perspective on series elastic actuators. 33:01.287 --> 33:03.387 So they have series elastic actuators. That 33:03.387 --> 33:05.887 idea is you put a spring, the output of a transmission 33:05.887 --> 33:07.787 between the transmission and the load, 33:07.787 --> 33:10.027 and sense the deflection of that spring, turn a 33:10.027 --> 33:12.127 force control problem into a position control 33:12.127 --> 33:14.707 problem, which is really much easier to solve. 33:14.787 --> 33:17.567 So that's on the upside. On the downside, 33:17.567 --> 33:20.627 they reduce torque bandwidth. So that's generally 33:20.627 --> 33:23.067 bad. And they add complexity and mass. 33:23.107 --> 33:26.187 So we wanted to build a design that could do both, 33:26.187 --> 33:28.107 that could either operate as a series elastic 33:28.107 --> 33:31.487 configuration or operate as a rigid actuator. So 33:31.487 --> 33:33.027 I'm going to talk to you about how we did that. 33:33.327 --> 33:35.327 To do that, we had to develop 33:35.327 --> 33:37.127 a new type of torsion spring. 33:37.707 --> 33:42.107 torsion springs you know rotational springs generally 33:42.107 --> 33:46.347 are very stiff and they're generally one part so 33:46.347 --> 33:49.057 they're one part usually they're sort of a serpentine 33:49.057 --> 33:52.627 spirally structure and that is that ends up being a 33:52.627 --> 33:55.507 very stiff spring because there are strain constraints 33:55.507 --> 33:58.087 on the beginning and ending of those spiral beams 33:58.087 --> 34:01.467 what we did was separate that into two parts so we have 34:01.467 --> 34:04.587 an inner sort of wheel and an outer set of spokes. 34:04.927 --> 34:09.287 And that made us get much lower stiffnesses 34:09.287 --> 34:12.367 than would be possible in the same volume. 34:12.967 --> 34:17.107 So low stiffness compact design. It fits inside 34:17.107 --> 34:20.087 the output pulley. So this is that belt drive, 34:20.087 --> 34:22.587 cross section of the knee. That's the belt 34:22.587 --> 34:24.727 drive. These springs fit in that output pulley. 34:24.727 --> 34:28.207 So it costs nothing in terms of volume. The 34:28.207 --> 34:30.867 springs have mass, but they don't add any volume. 34:32.157 --> 34:35.087 This is the spring. That's kind of me messing with 34:35.087 --> 34:37.167 one on my desk. You can kind of get a feel for 34:37.167 --> 34:41.007 it. So it's 24 radially spaced cantilever beams that 34:41.007 --> 34:43.427 interface with a gear -like camshaft. And when 34:43.427 --> 34:45.727 that camshaft is rotated, it deflects those beams. 34:46.067 --> 34:47.887 Cantilever beams are really easy to model. 34:47.887 --> 34:49.307 So we're able to get a really clear 34:49.307 --> 34:51.027 picture of the mechanics of these springs. 34:53.167 --> 34:55.607 So you have to select the series elastic stiffness 34:55.607 --> 34:58.167 upon assembly. So when you build this device, you can 34:58.167 --> 35:00.607 build it as a series elastic actuator if you want to. 35:01.627 --> 35:04.067 So the springs interface with this gear 35:04.067 --> 35:07.287 -like camshaft, which is this or this. 35:10.907 --> 35:11.747 Okay. 35:14.827 --> 35:16.387 This is sort of what I just mentioned. 35:17.707 --> 35:20.967 So these springs, this is what they look like. And 35:20.967 --> 35:24.267 we have this paper that describes four different 35:24.267 --> 35:26.427 versions of these springs that all are the same. 35:27.027 --> 35:29.127 These springs you're looking at here all have the 35:29.127 --> 35:31.707 same stiffness, even though they look very different. 35:31.887 --> 35:34.187 So, this is just showing those designs 35:34.187 --> 35:36.767 and their finite element analysis. 35:37.467 --> 35:40.247 We have a closed form expression based on strain 35:40.247 --> 35:42.847 energy that lets us determine or model the 35:42.847 --> 35:44.287 stiffness of these, and we can do that really 35:44.287 --> 35:48.067 well. So, we have accuracy within 5%, and we have 35:48.067 --> 35:51.527 at one stage, so not a series configuration, 35:51.667 --> 35:55.167 but in a single spring, we have almost no backlash, 35:55.437 --> 35:57.247 essentially no backlash in a two-part 35:57.247 --> 35:59.747 spring. So that's really good. It requires 35:59.747 --> 36:02.927 wire EDMing, but it works. It works great. 36:03.527 --> 36:05.527 Little to no hysteresis or backlash. 36:05.527 --> 36:07.007 They're extremely efficient. 36:07.967 --> 36:11.867 So then we took this idea and this analysis 36:11.867 --> 36:14.627 and put it into a software package. 36:15.227 --> 36:19.927 So this tool, which is available at that link, is a 36:19.927 --> 36:22.127 MATLAB application that you can just download and run. 36:22.127 --> 36:24.107 You actually don't even need MATLAB, but it will 36:24.107 --> 36:27.187 design the springs for you. So you have to give it things 36:27.187 --> 36:29.827 like material properties, your desired stiffness, 36:30.047 --> 36:37.307 and how many different beams do you want, but it 36:37.307 --> 36:40.047 will then generate the springs, and that's available. 36:43.027 --> 36:47.307 So then we began analyzing other springs in 36:47.307 --> 36:48.767 the robotics literature. That's what we're 36:48.767 --> 36:51.027 looking at here. It's other torsion springs kind 36:51.027 --> 36:52.707 of developed over the last 10 years or so. 36:52.987 --> 36:55.067 And so on this side, these are 36:55.067 --> 36:56.447 the springs that we published. 36:56.787 --> 36:58.927 These are the springs in the robotics literature. 36:59.847 --> 37:04.767 And you can see that this spring is able to have 37:04.767 --> 37:09.767 the highest energy density that we can find. So you 37:09.767 --> 37:12.627 can kind of tell by looking at the spring. It's very 37:12.827 --> 37:15.107 compact. It's dense. There's not much air in it. 37:16.247 --> 37:18.907 Okay, so we took that tool and 37:18.907 --> 37:20.087 we looked up these springs. 37:20.687 --> 37:22.487 and we redesigned the springs 37:22.487 --> 37:25.127 from the literature with a tool. 37:25.547 --> 37:27.907 So we found the springs in the robotics literature, 37:27.907 --> 37:30.227 redesigned them, and we were able to recreate 37:30.227 --> 37:32.277 all of the springs from the literature with 37:32.277 --> 37:34.647 smaller and lighter footprints in every case. 37:35.657 --> 37:37.797 So this is where I was saying you could see, 37:37.797 --> 37:39.527 like you can tell the energy density 37:39.527 --> 37:40.627 of these springs is going to be very 37:40.627 --> 37:42.767 high because there's almost no air. 37:42.767 --> 37:44.767 Like, look at that spring, you 37:44.767 --> 37:46.007 know, it has tons of air in it. 37:48.027 --> 37:48.827 Okay. 37:50.947 --> 37:53.107 I think this highlights the potential use of the 37:53.107 --> 37:54.867 spring. We can get back to the open source lag. 37:54.867 --> 37:57.347 So most people do not use it as a series elastic 37:57.347 --> 38:01.967 device yet. We haven't published the package for the 38:01.967 --> 38:04.607 Tor controller, which we will do hopefully soon. 38:05.847 --> 38:08.507 But OSL, open source lag, open access 38:08.507 --> 38:09.947 hardware and software platform, 38:10.487 --> 38:11.887 it's pretty cool. 38:12.547 --> 38:15.727 We use custom continuous integration steps 38:15.727 --> 38:19.647 to build these and deploy them using GitHub 38:19.687 --> 38:21.727 and Git Actions. GitHub Actions is awesome. 38:23.107 --> 38:25.937 One thing I think is really cool is we have this 38:25.937 --> 38:28.887 community of people using this hardware to study control. 38:28.947 --> 38:32.247 They can share their control files without sharing 38:32.247 --> 38:34.867 their source code. They can take their controller, 38:34.867 --> 38:37.347 compile it into a shared object, and share that. 38:37.347 --> 38:39.347 And so now we can test on each other's 38:39.347 --> 38:41.867 controllers without having to share source code. 38:41.867 --> 38:44.067 That's awesome. That's really valuable. 38:44.327 --> 38:47.867 We already have two researchers who have published 38:47.867 --> 38:50.167 their controllers, and they're available online. 38:50.547 --> 38:52.747 We have, like many researchers around 38:52.747 --> 38:54.507 the world, maybe about 20 or 30 38:54.507 --> 38:57.587 people using the leg around the world. 38:58.787 --> 39:00.527 This is just a little kind of run-through 39:00.527 --> 39:02.787 of the existing website. This website's about 39:02.787 --> 39:05.007 to change kind of in the next seven days. 39:05.347 --> 39:08.307 it's like it'll be it's much that we thought this 39:08.307 --> 39:10.527 is maybe too intimidating looking that's kind of where 39:10.527 --> 39:13.787 it's not user-friendly we're sort of like we feel 39:13.787 --> 39:16.267 like maybe it's a little too intimidating so we're 39:16.267 --> 39:18.107 about to make a change on the site that makes it 39:18.107 --> 39:22.387 warm and friendly but this is so that was the website 39:22.567 --> 39:26.287 you'll see a big change coming out soon and that 39:26.287 --> 39:28.527 this is just showing Kevin kind using the software. 39:30.287 --> 39:33.327 Okay, this is my last talk, 39:33.327 --> 39:36.427 my last presentation project. 39:37.027 --> 39:39.767 What I'd like to tell you now is about trying 39:39.827 --> 39:42.927 to build exoskeletons and building them well. 39:43.257 --> 39:45.127 And how do we know if we're doing that? 39:45.467 --> 39:49.307 So this is a collaborative project. 39:49.987 --> 39:52.727 Leo Medrano was a former PhD student. Craig 39:52.727 --> 39:54.247 Thomas was a former postdoc. Now he's at 39:54.247 --> 39:56.727 Texas A&M. And Drew Margolin is at Cornell. 39:58.127 --> 39:59.027 Okay. 40:01.627 --> 40:03.447 I think I'm just going to skip this. 40:03.747 --> 40:07.167 So the idea here is, how do we know we're 40:07.167 --> 40:08.927 building exoskeletons that do well for their 40:08.927 --> 40:11.207 user? How can we tell that we're doing that? 40:11.727 --> 40:15.307 How we assess or what the goal is of the exoskeleton 40:15.307 --> 40:18.167 factors into how we build it. So all of those 40:18.167 --> 40:20.687 decisions are sort of made up front. And the 40:20.687 --> 40:23.327 only thing people really use is metabolic rate. 40:23.427 --> 40:26.327 They're like, is there something else we can use? 40:26.767 --> 40:30.467 We have studied people's own perception to their 40:30.467 --> 40:34.007 metabolic rate. We have quantified that people 40:34.007 --> 40:36.847 cannot sense changes to their metabolic rate unless 40:36.847 --> 40:40.147 they're greater than 23%, which is a massive 40:40.147 --> 40:42.807 change. In fact, your metabolic rate is changing all 40:42.807 --> 40:45.327 the time, and you probably are not consciously 40:45.327 --> 40:48.327 aware of it. So we did that study, which is very 40:48.327 --> 40:51.667 interesting. But the idea here is, can we assess an 40:51.667 --> 40:54.647 Euclidean based on the value it adds during use. 40:56.057 --> 40:57.557 What do I mean by value? 40:58.307 --> 41:01.087 So in order to talk you through this experiment, 41:01.087 --> 41:02.377 I'm going to talk you through the protocol. 41:02.487 --> 41:05.887 So imagine you're walking on a treadmill, 41:06.767 --> 41:10.107 uphill 10 degrees, so pretty high, pretty fast. 41:10.107 --> 41:13.767 So 1.25 meters per second. This is no fun. 41:13.767 --> 41:15.927 This is a hard task. So 41:15.927 --> 41:17.387 imagine you're doing this task. 41:17.507 --> 41:18.807 It's no fun. 41:19.127 --> 41:20.887 Now imagine I say, 41:21.127 --> 41:23.727 how much money do I have to pay you for you 41:23.727 --> 41:25.827 to do this task for another two minutes? 41:26.767 --> 41:29.207 And then at the end of that two minutes, I ask you 41:29.207 --> 41:31.587 again, how much money would I have to pay you to do 41:31.587 --> 41:34.927 this for another two minutes? And we just keep going. 41:35.747 --> 41:37.227 You might imagine 41:37.847 --> 41:40.887 that as you get more fatigued with this task and 41:40.887 --> 41:43.817 annoyed, you're going to demand more money, more 41:43.817 --> 41:45.707 compensation. And in this project, you have to 41:45.707 --> 41:47.787 get the money afterwards. So this is all real. 41:48.117 --> 41:49.287 So that you might end up with 41:49.287 --> 41:50.267 something that looks like this. 41:50.987 --> 41:53.667 We'll call this price to walk. So with a dollar 41:53.667 --> 41:55.867 amount that you demand to walk for two minutes, 41:55.987 --> 41:58.427 and this is the time of the experiment. 41:58.567 --> 42:01.647 So you can see maybe as you go on, you're 42:01.987 --> 42:03.807 tired, you don't want to keep walking uphill, 42:03.967 --> 42:05.947 and you're not wearing an exoskeleton. 42:06.647 --> 42:08.467 Now let's imagine you re-complete this 42:08.467 --> 42:10.407 task, but I give you an exoskeleton that 42:10.407 --> 42:12.927 meaningfully assists you in the task. 42:13.247 --> 42:16.427 What we would expect to see is a decrease in this 42:16.427 --> 42:18.017 amount of money that you have to be compensated. 42:18.307 --> 42:21.367 If the exoskeleton is actually helping you and making 42:21.367 --> 42:24.867 the task easier, you should demand less money. 42:25.947 --> 42:29.387 The value of the exoskeleton is 42:29.387 --> 42:30.867 the difference in those curves. 42:31.087 --> 42:34.107 That's the value added by the exoskeleton in 42:34.107 --> 42:36.827 U.S. dollars for the task of walking uphill. 42:37.607 --> 42:42.447 So this is a kind of potentially very different 42:42.447 --> 42:45.347 way of looking at what these technologies can do. 42:46.847 --> 42:48.307 You might be wondering, 42:48.847 --> 42:51.307 well, what if they're just trying to, the participants 42:51.307 --> 42:53.627 just trying to make as much money as they can? What 42:53.627 --> 42:55.987 if they're not being honest and they're not telling us 42:55.987 --> 42:58.527 their actual amount, they're just trying to get money? 42:58.947 --> 43:01.207 So that's a legitimate question. 43:01.347 --> 43:03.247 And the way we got around that was by 43:03.247 --> 43:04.987 using something called the Vickrey Auction. 43:05.627 --> 43:09.027 So we need people to be honest about how 43:09.027 --> 43:11.947 much money they need or want to do this task. 43:12.147 --> 43:14.487 So we're using something called the Vickrey 43:14.487 --> 43:16.967 Auction, which is a sealed second price auction. 43:17.197 --> 43:19.447 This paper came out in the 1960s and 43:19.447 --> 43:21.047 won the Nobel Prize in Economics. 43:21.047 --> 43:24.487 And it is a method, a sealed second price auction, 43:24.487 --> 43:27.387 where the lowest bidder gets the second lowest bid. 43:27.507 --> 43:30.127 And this has been used to judge 43:30.127 --> 43:32.127 the value of abstract concepts. 43:32.567 --> 43:34.467 It's been used to assess the 43:34.467 --> 43:36.707 true value of non-GMO foods, 43:36.887 --> 43:39.107 Facebook usage, battery life. 43:39.127 --> 43:43.607 It's been used to assess the difficulty or 43:43.607 --> 43:45.687 consequence of drinking a bitter liquid. 43:45.687 --> 43:47.267 How much does that cost? 43:47.567 --> 43:49.487 That's kind of like wearing an exoskeleton. 43:53.387 --> 43:55.807 So now we have a sequential auction going 43:55.807 --> 43:57.507 on. Every two minutes is an auction. 43:58.067 --> 44:00.627 How much do you need to be paid? You have to bid on 44:00.627 --> 44:03.147 that. And if you win the auction, then you walk. If 44:03.147 --> 44:06.167 you lose it, you rest. So now we have a set of sequential 44:06.167 --> 44:08.927 auctions, which means we need multiple people. 44:10.077 --> 44:11.967 If there's auctions, there needs to be 44:11.967 --> 44:13.687 other people participating in this auction. 44:14.027 --> 44:15.507 We don't want to do that. 44:15.587 --> 44:19.127 So what we did was we used RoboBidders, so computerized 44:19.127 --> 44:21.927 bidding agents that were kind of trained on 44:21.927 --> 44:25.287 the behavior of human bids in pilot studies from 44:25.287 --> 44:27.727 three people. So we took their bids, we fit them to 44:27.727 --> 44:31.387 the similar mathematical model, and it increases if 44:31.387 --> 44:34.847 they walk, it decreases if they rest, sort of follows 44:34.847 --> 44:37.627 some similar profile that we could find from 44:37.627 --> 44:40.347 pilot data. So we don't need, this keeps us from needing 44:40.347 --> 44:43.347 extra subjects. There are three robot bidders. 44:44.187 --> 44:46.067 So yeah, robot bidders were based on the first 44:46.067 --> 44:48.087 -order models of human bids. The participants 44:48.087 --> 44:50.347 reached an equilibrium with their robot bidders. 44:50.347 --> 44:53.187 That's always true in sequential auctions. 44:54.487 --> 44:57.107 And so the goal here now is we're going 44:57.107 --> 44:58.887 to use this Vickrey auction protocol to 44:58.887 --> 45:00.907 measure the value added by an exoskeleton. 45:02.767 --> 45:04.907 This is neat. I'm excited to tell you about 45:04.907 --> 45:07.567 these results. So we had 16 subjects who completed 45:07.567 --> 45:09.867 this. So they are queried every two minutes, 45:10.447 --> 45:11.947 as I said, but they win, they 45:11.947 --> 45:13.827 walk. If they lose, they rest. 45:14.717 --> 45:16.527 So they don't win every time. They win like 45:16.527 --> 45:18.027 two-thirds or three -fourths of the time. 45:18.187 --> 45:19.927 Then we kind of bias that based 45:19.927 --> 45:21.507 on the settings in the experiment. 45:22.127 --> 45:23.947 So we tested them in three conditions. 45:23.947 --> 45:24.767 This is going to be important. 45:25.567 --> 45:27.887 No XO, so not wearing an XO at all. 45:28.167 --> 45:30.687 Wearing the XO, but powered off. 45:30.687 --> 45:32.747 And wearing the XO, but powered on. 45:33.007 --> 45:34.387 So three conditions. 45:34.487 --> 45:36.987 The controller we used was not our controller. 45:37.107 --> 45:39.667 It was a controller developed by Defy who makes 45:39.667 --> 45:43.347 the exoskeletons from around 2018, the controller. 45:43.527 --> 45:46.607 So we wanted it to be a proxy for the 45:46.607 --> 45:48.487 state of the art. We wanted it to be the 45:48.487 --> 45:50.127 best controller that we could obtain. 45:50.487 --> 45:54.127 I think that's probably Defy's, at least in 2018. 45:54.787 --> 45:57.147 What we end up doing is fit the bidding data with 45:57.147 --> 45:59.187 a first-order exponential fit to the log transform 45:59.187 --> 46:02.467 data, and then we calculate what we call this marginal 46:02.467 --> 46:05.647 value. So those integrals just contain those 46:05.647 --> 46:09.027 curves and calculate the area that I showed before. 46:09.547 --> 46:11.887 So once we, the bid actually, bids fit 46:11.887 --> 46:14.387 very closely to these exponential models, 46:14.387 --> 46:15.627 which I'll show you in just a minute. 46:18.847 --> 46:21.147 Yeah, so it's easy to obtain the 46:21.147 --> 46:22.747 coefficients through regression. 46:24.687 --> 46:27.707 Okay, so I'm also going to talk about this concept, 46:27.827 --> 46:32.407 this value MV, as a percent of the no XO value. 46:32.747 --> 46:36.987 and as a cost in dollars per hour, which we can get 46:36.987 --> 46:40.427 from the aggregated dollars demanded by the subjects. 46:41.837 --> 46:44.207 This is what the data look like. So here we're looking 46:44.207 --> 46:46.507 at these price to walk curves for a few subjects. 46:46.707 --> 46:48.947 These are subjects that received a positive 46:48.947 --> 46:51.367 value from the exoskeleton. So when they put on 46:51.367 --> 46:53.587 the exoskeleton, it made things easier for them. 46:54.047 --> 46:56.727 This is a subject that made things harder. 46:57.007 --> 46:59.287 They put on the exoskeleton and they wanted more 46:59.287 --> 47:02.847 money to walk. And then there are subjects who it 47:02.847 --> 47:06.257 doesn't really matter. They don't really care, or 47:06.467 --> 47:08.567 at least it doesn't provide any benefit or cost. 47:09.297 --> 47:10.947 Okay, so I'm going to talk 47:10.947 --> 47:11.817 you through some histograms. 47:12.507 --> 47:16.407 This is the marginal value for all 16 subjects. 47:16.747 --> 47:20.507 It has very slight positive value. 47:21.667 --> 47:26.877 The overall value was 4.2 % of the no-exo condition 47:26.877 --> 47:29.767 was a very large standard deviation. So the red 47:29.767 --> 47:34.267 people had a cost. The blue people got a value. So 47:34.267 --> 47:40.187 overall, the value of the exoskeleton was very low. 47:40.447 --> 47:45.987 There was about $3.40 that it's providing in 47:45.987 --> 47:50.527 that task. So bummer. Not my favorite result. But 47:50.527 --> 47:53.207 I think that's what We saw, we see a very large 47:53.347 --> 47:55.407 standard deviation, a very large spread in the 47:55.407 --> 47:58.247 data indicating the differences across people. 48:00.147 --> 48:01.787 But, you know, what if we 48:01.787 --> 48:02.897 look at the other conditions? 48:03.207 --> 48:06.127 You know, this is not my favorite result. 48:06.527 --> 48:09.267 But if we look at the other conditions, maybe we can 48:09.267 --> 48:11.807 kind of figure out what's going on behind this number. 48:12.987 --> 48:16.747 So, we also collected not just, you know, exo-powered 48:16.747 --> 48:20.467 off or exo-powered on and no exo. we also collected 48:20.467 --> 48:24.147 exo-worn but powered off. So that gives us the 48:24.147 --> 48:26.367 opportunity to do some different calculations here. 48:27.447 --> 48:30.907 So here we're going to compare in this marginal value 48:32.507 --> 48:35.607 wearing the exo -powered off with not 48:35.607 --> 48:38.187 wearing an exoskeleton. So this is the cost 48:38.187 --> 48:41.007 of wearing 1.5 kilograms on your feet. 48:41.337 --> 48:45.227 And so what we see is it's negative. So that's good. 48:45.477 --> 48:47.567 That's like sort of a check mark for us. 48:47.567 --> 48:49.327 We should see a negative value when people 48:49.327 --> 48:51.107 put weight on their feet, you would think. 48:51.767 --> 48:55.327 And what we see, though, is a huge negative cost. 48:55.527 --> 49:00.487 So it swings to about minus $19 an hour 49:00.847 --> 49:04.867 to wear the added mass on your feet. 49:05.267 --> 49:08.147 Now if we compare exo -powered on to exo-powered 49:08.147 --> 49:10.327 off, we can get the value of just the assistance. 49:11.157 --> 49:12.327 That's what we did. 49:12.667 --> 49:16.087 And what we get is that the value of the assistance 49:16.087 --> 49:20.167 is really something like $20 an hour, but that 49:20.167 --> 49:24.327 value is consumed by wearing the mass on their feet. 49:25.887 --> 49:29.587 So despite the idea that exoskeleton assistance 49:29.587 --> 49:31.967 can be very, very valuable to people, 49:33.167 --> 49:35.947 when you add the mass to their feet, it's not so much. 49:35.947 --> 49:40.307 So that's both very interesting, a little bit of a 49:40.307 --> 49:44.427 bummer, but also there's some encouragement there too. 49:46.277 --> 49:48.447 All right, these are takeaways. 49:49.177 --> 49:51.507 User preference can be used as a meta 49:51.507 --> 49:56.627 -criterion to optimize technologies, to enable 49:56.627 --> 49:58.707 wares to have the best possible experience. 49:59.917 --> 50:01.787 When we used the VISPA foot, 50:02.887 --> 50:07.427 users, the only thing we saw was that users preferred 50:07.427 --> 50:10.147 to increase kinematic symmetry with their prosthesis. 50:10.367 --> 50:12.347 Otherwise, they had very diverse 50:12.347 --> 50:14.067 yet consistent preferences. 50:14.927 --> 50:18.667 New actuator design architectures may enable powered 50:18.667 --> 50:21.687 apparel or closer, more form-fitting exoskeletons, 50:21.687 --> 50:23.587 exoskeletons that don't look like exoskeletons. 50:24.177 --> 50:25.987 Open source lag is a platform to 50:25.987 --> 50:27.707 facilitate controls investigations. 50:28.147 --> 50:30.187 Please check it out if you're interested. 50:30.767 --> 50:33.547 And then new methods kind of from behavioral 50:33.547 --> 50:36.907 economics might provide us some information that 50:36.907 --> 50:39.507 we hadn't had access to around the benefit of 50:39.507 --> 50:41.287 these technologies or how to make them useful. 50:41.437 --> 50:43.207 So this is what I started with. 50:43.727 --> 50:45.707 Hopefully you kind of see a little 50:45.707 --> 50:47.357 bit about kind of how we do this. 50:48.087 --> 50:51.027 But thank you very much for your time. 50:51.027 --> 50:53.247 It's been a blast already. This is my 50:53.247 --> 50:55.967 team at Michigan. Come visit us sometime. 50:56.327 --> 50:58.487 And I think I can take some questions. 51:08.357 --> 51:09.017 Awesome. 51:09.177 --> 51:10.057 Yes, 51:21.967 --> 51:23.267 what exactly is it? 51:26.367 --> 51:29.307 I think the question is what exactly do I mean when 51:29.307 --> 51:31.587 I say kinematic symmetry with their prosthesis, 51:31.947 --> 51:32.567 right? 51:32.647 --> 51:35.727 Okay, so what that means, so every step they're 51:35.727 --> 51:38.577 making motion with their intact ankle and their 51:38.577 --> 51:41.347 prosthetic ankle. So every step, you know, 51:41.347 --> 51:46.187 those signals would progress. And so what this is 51:46.187 --> 51:48.127 saying is when they were at their preferred 51:48.127 --> 51:51.247 stiffness, those trajectories were most similar. 51:51.447 --> 51:56.227 And so somehow the trajectory of their biological ankle 51:56.227 --> 51:59.127 was closer to the trajectory of their prosthetic 51:59.127 --> 52:01.907 ankle that they have no sensation of. So that's why it's 52:01.907 --> 52:04.167 kind of like a little surprising, but interesting. 52:04.387 --> 52:07.567 But it really has to do with the exact motion 52:07.567 --> 52:10.507 of their ankle joint and the prosthetic ankle 52:10.507 --> 52:13.267 joint and how similar those profiles are in a 52:13.267 --> 52:15.427 step. And that's exactly what it's saying. 52:15.427 --> 52:17.847 They're very similar. They're maximally similar. 52:18.627 --> 52:19.467 Yeah. 52:19.707 --> 52:20.387 Yes. 52:25.117 --> 52:27.077 Well, I still want in that order. 52:27.077 --> 52:29.337 Or did you see any trend depending on 52:29.337 --> 52:31.397 which order you gave it to participants? 52:31.797 --> 52:33.897 That's a good question. We randomized it because 52:33.897 --> 52:35.537 we do think there was an effective order. 52:35.597 --> 52:37.237 So we randomized it. 52:37.237 --> 52:39.577 I can't really tell you why I think that. 52:39.897 --> 52:42.937 Like, it takes time for people to get used to 52:42.937 --> 52:45.857 the assistance. So even in all of our experiments, 52:45.857 --> 52:48.037 we would have them come in for a day where 52:48.037 --> 52:50.397 they just try it and just kind of get used to 52:50.397 --> 52:53.847 wearing it so like that that part is important but 52:55.137 --> 52:57.877 that's pretty much how we dealt with deal with that 52:57.877 --> 52:59.777 we ran you know we random we expect there's an 52:59.777 --> 53:03.677 order effect so we randomize and then we you know 53:03.677 --> 53:06.877 try to steer around it but it's a good question 53:06.877 --> 53:13.247 yeah so the activator is really cool thank you yeah 53:18.287 --> 53:22.427 how do you like turn that axis this. So we have one 53:22.427 --> 53:24.947 option. So this is like if it's rotating this way, but 53:24.947 --> 53:27.467 yet we want to apply torques in the sagittal plane, 53:27.467 --> 53:30.347 somehow there has to be some rotation there. So we 53:30.347 --> 53:34.047 have one version of that now, which is a cable-based. 53:34.067 --> 53:36.647 So it has a cable that goes around the hoop, and that 53:36.647 --> 53:41.727 cable just translates it to in the sagittal plane. 53:43.307 --> 53:45.727 One thing, though, so this is like that we 53:45.727 --> 53:48.557 have on the larger version of this hoop actuator. 53:48.587 --> 53:51.507 And so what's interesting about this concept 53:51.507 --> 53:55.727 is the hoop actuator is large, yet we want 53:55.727 --> 53:58.227 a transmission ratio that amplifies torque. 53:58.247 --> 54:03.547 So that is sort of ill -posed. So we have to first go 54:03.547 --> 54:05.987 have a negative transmission, essentially the opposite 54:05.987 --> 54:09.287 direction, and then two stages that invert that. And so 54:09.287 --> 54:11.647 that's not ideal. We haven't quite figured out how to 54:11.647 --> 54:15.707 do that better. But currently, that's a complication. 54:33.067 --> 54:36.247 That's a great question. 54:36.267 --> 54:39.907 My bet is, so another thing that I think that we 54:39.907 --> 54:43.327 have to try to keep from impacting our experiments 54:43.327 --> 54:46.747 is that subjects just like to be subjects. 54:46.777 --> 54:49.047 They like to come in and they like to do good, they 54:49.047 --> 54:52.647 like us to be excited, so it's really hard to have 54:52.647 --> 54:54.827 the subjects not automatically have a really positive 54:54.827 --> 54:59.247 bias so that's the like the trick is to kind of 54:59.247 --> 55:01.847 prevent that so even if we did it with ankle weights 55:01.847 --> 55:03.687 so that's if we i'm thinking like what if we did 55:03.687 --> 55:06.087 the 1.5 kilogram ankle weights would it would it 55:06.087 --> 55:09.507 produce a similar like trajectory of curves and i'm 55:09.507 --> 55:12.567 not sure because of this concept because the subjects 55:12.567 --> 55:15.777 love to come like they they enjoy being research 55:15.777 --> 55:17.797 subjects and working with us and wearing exoskeletons 55:17.797 --> 55:21.107 like it's fun and exciting So they're happy and 55:21.627 --> 55:24.407 it's really hard to get them to narrow to their 55:24.407 --> 55:26.757 preference. They're usually just happy with anything. 55:28.827 --> 55:31.087 So, yeah, yeah, I think I got it. 55:31.207 --> 55:35.747 Yeah, yeah. 55:37.267 --> 55:40.167 It went up to like 75 degrees Celsius. 55:40.167 --> 55:42.207 Yes. Are there any plans on how you're 55:42.207 --> 55:43.927 going to make that actually wearable? 55:44.047 --> 55:48.347 Yeah, well, so that test is really just fitting 55:48.347 --> 55:50.827 those thermal properties like that's how we identify 55:50.827 --> 55:54.607 kind of how hot it's going to get in use so that 55:54.607 --> 55:57.287 test is not is not really meant to be like a test that 55:57.287 --> 56:00.127 represents use that's a test that that generates 56:00.127 --> 56:05.007 the values the coefficients but certainly like heat 56:05.007 --> 56:08.407 is an issue my personal take on heat is that we 56:08.407 --> 56:10.767 want these systems to get hot and if they're not 56:10.767 --> 56:13.147 getting hot we're carrying around a bunch of mass 56:13.147 --> 56:15.907 actuator mass that we're not using so like there's a i 56:15.907 --> 56:17.407 think there's a balance here like we want to burn 56:17.407 --> 56:21.367 anybody for sure but in my opinion if an actuator 56:21.367 --> 56:24.147 is staying cool it's much too large you know so it's 56:24.147 --> 56:27.487 like this is this trade -off but it's a good point 56:31.227 --> 56:32.667 okay yeah 56:35.867 --> 56:36.787 yep 56:55.847 --> 57:00.907 what controller this is um i think it's just 57:00.907 --> 57:02.547 the standard state machine that's been tuned 57:02.707 --> 57:05.427 I don't think the best controller, the walking 57:05.427 --> 57:07.667 controller we have is Bobby's, Bobby Gregg's, which 57:07.667 --> 57:09.967 is one of the ones you can download. That's 57:09.967 --> 57:13.267 why I was trying to see if it has an IMU right 57:13.267 --> 57:16.977 there, which I can't really tell. But if it's Bobby's 57:16.977 --> 57:19.267 controller, it has an IMU right in the front. 57:19.267 --> 57:21.687 So it's either Bobby's controller or it's 57:21.687 --> 57:23.967 just we tuned the state machine to look nice. 57:25.287 --> 57:26.127 Yeah. 57:26.587 --> 57:29.507 Yeah. We mostly look to other people 57:29.507 --> 57:32.147 for the control, like in general. 57:32.267 --> 57:38.487 people like Aaron okay one more 57:38.487 --> 57:41.867 yes what yeah yeah yeah yeah 57:57.857 --> 57:58.577 yeah 58:17.417 --> 58:20.737 very different right people have very different 58:20.737 --> 58:24.337 sort of static values of their time which comes up 58:24.337 --> 58:27.577 in that experiment we have to navigate that um 58:27.577 --> 58:30.397 we would I would say we so for like how do we know 58:30.397 --> 58:32.837 they're being honest the way we are doing that 58:32.837 --> 58:34.737 is with this victory auction which is essentially 58:35.157 --> 58:38.227 being honest is the optimal strategy so 58:38.227 --> 58:40.297 that's as kind of as far as it goes is that 58:40.597 --> 58:44.257 that in in kind of behavioral economics that's what's 58:44.257 --> 58:48.137 been shown is that the optimal strategy is to provide 58:48.137 --> 58:50.317 a truthful amount there's a there's a rationale 58:50.317 --> 58:53.037 for that but it's kind of hard to explain but but that's 58:53.037 --> 58:55.257 that's as far as we take it so we don't really you 58:55.257 --> 58:57.187 know we don't really have any way to know that but 58:57.187 --> 59:00.047 if we see like sometimes you can see bidding behavior 59:00.047 --> 59:03.137 that is clearly indicates that they're doing something 59:03.137 --> 59:05.397 different. They're taking a different strategy. 59:05.637 --> 59:06.777 And that, 59:06.977 --> 59:09.177 like, as long as it's acceptable, they 59:09.177 --> 59:11.737 aren't clearly trying to make a ton of money, 59:11.797 --> 59:14.597 like, we leave it, you know? That's their strategy, 59:14.597 --> 59:17.017 it's their strategy. So sometimes it wouldn't 59:17.017 --> 59:19.997 exactly fit that exponential curve in some way. 59:19.997 --> 59:22.197 Sometimes they, like, stair-step a little bit, 59:22.197 --> 59:25.757 but they all fit, like, kind of relatively well. 59:25.757 --> 59:28.377 They all kind of tend to increase over time. 59:28.837 --> 59:30.597 Is that a time limit? 59:30.837 --> 59:33.277 balance that they knew beforehand. 59:33.277 --> 59:35.217 Yes, there's a time limit. 59:37.697 --> 59:39.737 That's right. It's like 45, approximately 59:39.737 --> 59:41.617 45 minutes, about 30 minutes of walking. 59:43.257 --> 59:44.157 Yep. 59:44.677 --> 59:45.777 All right. 59:47.907 --> 59:48.977 Thank you, guys. 59:48.997 --> 59:49.437 This 59:54.997 --> 59:57.177 week is machine learning seminar. 59:57.177 --> 59:58.547 See you again in two weeks. 59:59.037 --> 01:00:00.217 Well, I'll tell that