Ekso Bionics Holdings, Inc. (EKSO) Earnings Call Transcript & Summary

Welcome to AI for Good, the leading action-oriented, global and inclusive United Nations platform on AI, organized by ITU in partnership with 40 UN sister organizations and co-convened with Switzerland. The goal of AI for Good is to identify practical applications of AI to advance the United Nations Sustainable Development Goals and scale those solutions for global impact. In today's session, we're counting on you to use the live video wall feature to ask questions and post comments to help create an engaging discussion. We encourage you to stay until the end to chat, connect, ask questions and network with our distinguished panelists and world-class AI experts in the neural network. It is now time to kick off the session and welcome our first speaker. The floor is yours.

Thank you, Anna. Hello, everyone. Good morning, good afternoon, good evening from me as well. I am Guillem Martinez Roura from the International Telecommunication Union, and I would like to welcome you all for joining us for today's AI for Good online event on AI-powered Exoskeletons Revolutionizing Rehabilitation and Mobility. This session is part of the AI for Good Robotics programming track, exploring how AI-powered robots are helping to unlock our human potential and advance the United Nations Sustainable Development Goals. We have a distinguished set of panelists today. But as always, we are counting on you, the participants, to help create a very interactive discussion and session. Without further ado, I would like to introduce our moderator today. His name is Olivier Lambercy, and he is the Adjunct Professor at the Department of Health Sciences and Technology at ETH Zurich, and he is the Co-Director of the Rehabilitation Engineering Lab. Olivier, welcome, and the floor is all yours. Thank you very much.

Thank you very much, Guillem. A very warm welcome to everyone also from my side to this exciting webinar on the topics of AI-powered Exoskeletons Revolutionizing Rehabilitation and Mobility. My name is Olivier Lambercy, and it's my great pleasure to moderate this webinar for you today. So we're going to talk about robotics. We're going to talk about exoskeleton, how we design them, how we control them and how these devices can make a difference in the quality of life of the many people living with severe mobility impairment. We're going to discuss as well how artificial intelligence can play a role in how we develop these devices. So will we see tomorrow assistive exoskeletons everywhere around us, helping people with disability? Are we there yet? So let's see what our speaker thinks about that. So we have an exceptional lineup today of experts from the field, both from academia and also from industry, and they will give you a glimpse of their research on powered exoskeleton. They will each have a 10-minute presentation. And after the floor talks, we will have some time for discussion and for your questions. So please feel free to interact with us, write your question in the neural network interface, and then we'll bring them up to our speaker later on. So, without further ado, it's my great pleasure to introduce our first speaker. That will be Dr. Mohamed Bouri. He is the Director of the REHAssist Research Group at Ecole Polytechnique Federale de Lausanne, EPFL, in Switzerland. And his talk today will be titled Lower Limb Exoskeletons: from Designs to Daily Living and Rehabilitation Challenges - Use Cases. Mohamed, I'll let you share your screen. I'm really looking forward to your talk.

Thank you, Olivier. Hello, everyone. I'm really pleased actually to be with you today to share our developments and some of our achievements at the group of REHAssist at EPFL. I will speak about some challenges in design of lower limb and design aspects of lower limb exoskeletons, and I will also share with you some design rules. Okay. Let's start with where we come from. Our group actually has a lot of heritage in the development of powered orthosis. We have already been involved in the development of this lower limb orthosis. At this moment, we did call them orthosis since they are actuated for this one at the hip and the knee and the ankle, providing the sagittal movement of the limb. Then we developed this WalkTrainer, which is probably one of the precursor of our developments in lower limb exoskeletons. Then we moved to the development of this hip orthosis, which is called HiBSO. HiBSO is hip ball-screw-driven orthosis that we have developed. And at this moment, actually, the weight was very high. So, we didn't achieve that much results, but we have -- we did a lot of progress in the development of these exoskeletons. Then what I can share with you is another development, which is very interesting that we did with the University of Duke, with the objective to manage the movement of rhesus macaques, which have been implanted with brain electrodes. So, afterwards, in 2014, we had this idea actually to develop lower limb exoskeleton to assist gait. And at this moment, we have been targeting teenagers. So, we have decided actually to be driven by simplicity, modularity and versatility and to go very rapidly to the development of this lower limb exoskeleton, so simplicity, modularity and versatility with respect to these 3 points, the point related to the design, the mechanical design, related to the control, as well as related to the versatility with respect to the applications. We decided at that moment already to have only actuated hip and knee. This is after 18 months actually that we did develop the first -- no, after 10 months, developing the first leg, which is actuated at the hip and knee. It was the first demonstrator just to demonstrate the aspect of moving and actuating the hip and the knee. So, what you can see is that the exoskeleton, TWIICE, which has been driven by a very smart team, Tristan Vouga, Jemina Fasola and Romain Baud, and also Julien Pache, to have this success and to develop this very smart and very lightweight exoskeleton. This has been demonstrated at the first Cybathlon already in 2016. And in 2018, we have made another progress. Here, actually, I shared with you the development that we did in 2016, and you see that it's very simple because this is the idea is to be driven by simplicity, as I said. Two segments, the hip -- one of the hip and one of the knee, be -- having very simple control triggered by -- for this one, by buttons just to trigger the movements of the lower limbs. Then in 2018, we did a very nice, let's say, progress by progressing in different aspects, progressing first at the adjustability to be adjustable to -- up to people of 1.80 meter of height, up to people weighting of 100 kilograms, having an exoskeleton, which is very lightweight, weighting less -- about, let's say, less -- yes, less than 17 kilograms or 100 Newton meter per joint, having this modularity, the ease of transfer, as you can see there, and also the fact that it's something which is very, very modular and versatile with respect to these different applications. And I will share also this with you. So, here, you see Silke Pan wearing the device. You have -- you saw also the ease of transfer. The control -- we kept the control in the same way, okay? And with the -- also the same amount of autonomy, which is about 3 hours, depending on the task for sure. Then let me present to you something very interesting. This is the INSPIIRE that has been presented at ICORR 2022, which is a passive exoskeleton. It has been developed passive for healthy users without any activation. By the way, INSPIIRE, let's say, is an instrument. It's an instrument to inspire or to learn from healthy human how we do work and how we do manage the balance. And we did profit by doing this. So we did learn actually how healthy users are managing the sagittal balance and in order actually to transpose or to develop this bio-inspired control strategy to manage the sagittal balance. So here, I shared it with you. So we see TWIICE actually managing and having this management of the balance in the sagittal plane without any crutches. And here, we see Silke Pan, who is T10 paraplegic -- T10 complete paraplegic, and you see that she manages the balance without actually managing her upper body and without having this capability to manage the lower body. Another -- let me share with you another challenge, which is related to backdrivability because as well as we did reach this 100 Newton meter, we kept actually the challenge to have this backdrivability with respect to -- with the objective to have this partial assistance for stroke patients who are capable to walk with one limb, while the other one is impaired. Here actually is the approach that we have developed to develop this rehabilitation strategy based on mirroring, or what we call also echoing, in which we echo the non-affected limb on the affected limb. Here actually, the exoskeleton is passive, which means that there is no assistance at all, while here, actually the echoing is active. And you see actually that the walk assistance works very well. And hopefully, we will translate this to clinics. Let me share with you other kinds of use cases as promised. And here, you see actually the clipping, the paradigm in which instead of having the devices -- the environment which adapts to the people with impairment, here, we see that the exoskeleton is capable to adapt to outdoor and to adapt to the possibility actually to have the exoskeleton outdoor. Another challenge also related to the outdoor and daily living applications is the dream of Martin Loos, who did contact us at the end, with this dream to do ski touring. So we adopted actually the exoskeleton, and we did prove actually this possibility to have this implementation of ski touring. I have some issue with my videos. Let me just get out from the -- okay, I'll do it again. Okay. So you can see actually the possibility that we do this ski touring. That's another -- let me share with you another exoskeleton with other challenges. So we have developed this eWalk. eWalk is a lower limb exoskeleton that we developed probably 2, 3 years ago with the objective to assist people with -- healthy people and particularly healthy users and to promote physical activity and to promote outdoor activities. So here, I share with you this approach in which we do learn actually preferences, okay? We do learn preferences. And we have, let's say, with -- using this pair-wise approach comparison with -- by using actually different profiles by doing the parameterization of the torque profile to assist people, we had actually reached the preferred profiles to assist working, which means here actually not that AI is using to define the preferred profiles, torque profiles to assist working, but we have also the scoring from which actually we are far, let's say, closest or far from this preferred working. Let me skip this and share with you some challenges or some of our dreams. The idea actually next is just -- before ending my presentations is that we reach -- by developing more and more these devices, by developing more and more the control approaches that by promoting the accessibility of these devices that we reach the way that the people -- these devices will not be called any more machines and not robots, that they are called devices, okay? There are tools and instruments, let's say, equipment, daily living equipment, and then, also one of the -- and then also by improving the wearability. And I enumerate some challenges that are related to working on the personalization and adaptation of the tasks -- to adaptation of the task to the terrain. And finally, one of the objectives that we do define in my research group is to help by using these devices to balance the -- and to manage the balance, while actually having still the lightweight exoskeletons. And thank you so much. I'm done with my presentation. Thank you, Olivier. The floor is yours.

Thank you so much, Mohamed, and very, very impressive to see all the fantastic devices your group has been developing, so congratulations. I will keep the questions for the end. So, we'll move directly to our second speaker today. That will be Dr. Hao Su, who is an Associate Professor of Mechanical and Aerospace Engineering at North Carolina State University in the United States. And his talk will be about small and smart soft wearable robots for everyone via high-torque motors and artificial intelligence. The floor is yours.

Yes. It's great pleasure to be here today. Thank you for having me. I would like to share about our work, how to leverage high-torque motors and AI to make soft robots -- soft wearable robots, smaller, smarter for everyone, for everyday applications. Our development goal falls in 2 dimensions. The first is, we want to make our robots through design -- hardware design to be more lightweight compliant to have a really comfortable human-robot or human-device interaction, so we can help the robot -- leverage the robot to reduce energetics during human locomotion or during human manipulation. If you look at the other dimension is, we want to leverage the control, in particular, learning-based control methods to make a robot really versatile. So it not only can help walking, but also can help other activities like stair climbing, even running in community or home settings, not only limited to lab settings. If we consider the anatomy of any robotic exoskeleton, any powered exoskeleton, it's typically composed of motors. Many of them have low torque. So we need to use high ratio gears to amplify the torque. And usually, there are some kind of like rigid transmission, rigid linkages. In our lab, we want to leverage high-torque motor, some custom high-torque motor developed in my lab and in concert with a soft transmission mechanism, including cable, fabric, textile or elastomer to make the physical interface with the human more comfortable and more natural. If we think about the first actuation approach, primarily about using electric motor for compliant robots, our design objective is to make robot compliant, high bandwidth. This means faster response and also highly efficient, so the robot can last longer with a single battery, and also even to reduce the cost of overall device. If we think about the traditional actuation method and also kind of -- some kind of a recent actuation method called the series electric actuator, they all somehow kind of have the mediocre bandwidth and not very compliant. And the overall efficiency is also mediocre. So through this quasi-direct drive actuation, basically the custom motor design that have high torque capacity, and we can use low-ratio gears to overall maximize all those design parameters. Yes, to achieve this kind of design objective, of course, there's no free lunch. So we have been working on physics-informed motor design. And to make the motors to have a smaller dimension means we can reduce -- we are able to reduce the motor diameter and also significantly reduce the thickness of those motors. So when we put those motors and low-ratio gears together and actuator, we can design this kind of really lightweight and ultra-compact exoskeleton that can not only work in the lab setting, but truly kind of really compact and become kind of consumer electronics, and Mohamed said, is kind of becoming a device instead of robot or machine, right? So, our latest version is about 2.5 kilogram for bilateral and bilateral configuration. Another aspect is, we want the robots not only to help people with neurological impairments, but help people with musculoskeletal impairments like for people with knee osteoarthritis. In collaboration with Dr. Ann Spungen and the U.S. Department of Veteran Affairs, we were able to test like 6 subjects to significantly reduce pain using a knee exoskeleton with our lightweight compliant exoskeleton for the knee joint. Using the similar actuation approach, we are also able to design a very lightweight, compact, portable shoulder exosuit. This is cable-driven, and we are able to significantly reduce the muscle efforts with this upper limb shoulder exosuit. In addition to this kind of compliant robot, but it's kind of sometimes people call this rigid but compliant robots, we also work on softer robots, but also leveraging this high-torque motor solution. We use a similar high-torque motor but maybe a little bit smaller, and we also developed a custom magnetic pump. So, these kind of robots can -- this kind of actuator can provide actuation for fluid-driven exosuits. The overall idea is similar to the compliant actuation. But overall, we want this robot to be fully portable and significantly enhance its efficiency, but also generate high force that's required for exoskeleton applications. Using this kind of magnetic pump, we are able to develop untethered hydraulic exosuit for the lower limb. We also have a fluid-driven exosuit for the lower limb. And we want to get rid of this tethered fluid-driven actuation platform by leveraging this high-torque motor and the custom magnetic pump. In addition to the lower limb, we also work on the upper limb exosuit modular design for the elbow joint and the shoulder joint. It's the same attrition, but it's able to significantly save -- reduce human effort to make human manipulation easier for both able-bodied individuals and also people with disabilities like AF population. In addition to make robots smarter and lighter, we also want to make robots really kind of more efficient, also smarter and more intuitive to work with the human. AI is very powerful, but the problem and the challenge with AI is that usually it requires massive data. And in our solution, we have been developing this model-based and data-driven approach. We call this experiment-free solution. By using -- by leveraging simulation, we are able to generate a significant amount of data in simulation environment to really lend the control policy. And we can also deploy this kind of control policy in a physical robot, and it can immediately help a human -- make a human walking easier. And also, it can not only assist the walking, but it can also assist running and stair climbing and also save the largest amount of energetics during all 3 kinds of locomotion activities. And again, this is based on some kind of simple but efficient neural network architecture, and it's very -- have been proven to be this kind of experiment-free approach is feasible and also beneficial for human locomotion in the real-world scenarios. We were very fortunate to have this work published in Nature this summer. And this is the original cover art we designed. We want to explain that this simulation and also learning, they can work together and very useful. This is actual Nature cover, and our paper is actually here. So you can see they are very different. But we feel very fortunate to demonstrate that this approach is feasible, so we can let human use robots in a very intuitive and efficient manner. So, if we think about where we are in terms of assistive robot or wearable robots, I think we are -- if we use a smartphone as an analogy, I think we are still in the early stage of this kind of research, and we think there will be rapid development and also deployment and even adoption of exoskeleton, not only for -- to improve mobility for people with impairments, but also maybe for everyone like for other consumer electronics for outdoor activity assistance. And I think we are very close to this becoming a reality. Thank you.

Fantastic. Thank you so much, Dr. Su, for this super inspiring presentation, impressive hardware and also first glimpse into how AI can be used in exoskeleton development. And I think we will directly continue on the similar topic with our next presentation. It's my great pleasure to welcome Dr. Shuzhen Luo, and she is an Assistant Professor in the Department of Mechanical Engineering at Embry-Riddle Aeronautical University in the United States. Shuzhen, you're here. You can share your slides, please.

Okay. Let me share my screen.

And turn on the video. So her talk will be on experiment-free exoskeleton assistance via learning in simulation. Thank you very much. The floor is yours.

Thank you so much, Olivier. Thank you so much for the introduction. It's great to be here today. I can hear some of the recent studies, interesting results already in this session. My research goal is to investigate the learning in simulation and realize the experiment-free exoskeleton assistance in the real-world settings. We aim to simulate human-robot interaction and directly use it to improve the physical mobility. We have been working on the neuromechanical simulation and leveraging AI computing-based technology to analyze the human-robot interaction. We also invite -- working on the machine learning-based autonomous control for the intelligent adaptive system to improve the physical mobility in the real world. We can also investigate the human biomechanics effects on the human body to the robotic systems. My research goal is to investigate the closed-loop interaction simulation, which can reduce the real human intensive testing as much as possible. The robot will first perceive the human kinematic states to detect human motion intention. And then, the robot will start to generate optimal assistance torque profiles, which is synergistic with human. So in these closed-loop human-robot interactions, it has some many aspects we need to consider to simulate in this interaction simulation. And there are several challenges, including from both the human and the robot. So, for the human body, it's very -- the human has a very highly dynamic system. It is the autonomous muscle-controlled human agent. We can provide the active interaction or response to the robot assistance. And there are some variations in terms of human kinematics, human motion state and muscle response in the different individuals. So, how to incorporate -- I mean, autonomous of the human agent and with the autonomous ability to perform different activities in the simulation, which is also like one major challenge in this field, and for the robots, how to incorporate the human feedback like motion response or the muscle response to the control loop, and the robot can quickly identify the human motion intention and generate optimal torque assistance profiles in the simulation. So they are the major challenges in this field. And our solution is to investigate the dynamic-based data-driven reinforcement learning [indiscernible]. For the human -- for this autonomous muscle control system, we designed the 2 neural networks, including the motion imitation network and the muscle coordination network. So, for the first motion imitation network, the human agent, we learned how to perform the different activities like walking, running and stair climbing, given the current human kinematic states. So the input is human kinematic states of the lower body and upper body. The output is the reference joint torque to support the different activities. So, since the human's muscle-controlled system -- so we designed the second muscle coordination network, which can generate the corresponding muscle activations to support -- to track the reference joint torques. To have the most stable moment in the training process, we designed the reward function by using the center pressure, which can be controlled in a stable region. So, in this way, we can like -- by designing the motion imitation network and the muscle coordination network, we can have autonomous -- we can have a 3D musculoskeletal human agent to perform the multiple activities with a good balance and stability. So the input is only full body human kinematic states. The output is the muscle activations for all the -- including the lower body and upper body. The second, for the robot, we want to design the controller to generate optimal torque assistance that can reduce the human muscle effort. We want to have the end-to-end -- we want to design the neural network-based controller, which take the human kinematic states as the input and direct output, the optimal joint torque assistance to reduce the human muscle effort. And we introduced the end-to-end control mechanism as they directly match the human kinematic states to the joint torque assistance. We want to have the adaptive system make the control process more intelligent, more straightforward without any intermediate parameters tuning or the human involved intensive testing in the control process in the real-world settings. That is the neural network-based exoskeleton controller. So the next thing, since we want to simulate this closed-loop human-exoskeleton interaction process, that means we need to consider the human response to the -- we need to incorporate the human response to the control loop, and the controller will generate optimal assistance to the human body and regenerate the new state -- new reward function. So, by considering this, we do jointly training these 3 networks to simulate this closed-loop interaction process. When the 3 neural networks all achieve the maximum reward, the exoskeleton controller has already learned how to autonomously reduce muscle effort to generate the optimal torque assistance and quickly identify the motion -- or the human motion intention. That is the -- yes, that is the data-driven and dynamics-based data-driven learning framework in this work. This is the learning process of this neural network-based exoskeleton controller. Almost like we -- after taking around 6 hours, the neural network starts to converge. And after taking around 8 hours, it can achieve a very good convergence. So the exoskeleton already learned how to identify the human motion intention by using the human -- by using the motion sensor -- by using the motion states, like kinematic states, including the joint angle and the joint angular velocities. We only train once -- only train around 8 hours once. The robot controller can learn how to assist -- how to reduce muscle effort, improve the mobility and also for the different individuals. The second challenge after is the simulation-to-reality gap. When we have trained the exoskeleton control policy in the robot interaction simulation, we want to transfer this control policy in the real world, the hardware. But there are some unique challenges, including the model uncertainties from both the human body and the robot modeling in the simulation, and due to some human variations in the kinematic states and the motion state and the different muscle conditions in the different individuals. The last one, if the human generates some unexpected behavior, so the control policy in the simulation often doesn't work in real. So how to overcome the sim-to-real gap for this kind of the human-robot interaction system is still one major challenge in this AI-powered, AI-controlled robotics field. Now, in our work, we consider -- we investigated the muscle dynamic randomization to incorporate the variations of the human muscle. In addition to considering the dynamic randomization for the robot, we try to randomize the motor strength, the size, the center of mass, the observation latency or the control delay or inertia for the robot model to simulate the robot model uncertainties. In addition to considering the robot dynamic randomization, we also take into account the muscle -- the variations of the muscle -- of the human muscle in different individuals. We try to randomize the maximum muscle isometric forces in the muscle model to incorporate the human variations in this closed-loop interaction process. So, after that, we do -- we can start to do the simultaneous training of the 3 neural networks. And then, after we incorporated the muscle dynamic randomization in this closed-loop interaction process, the exoskeleton controller, the policy learned from the simulation can be directly transferred or deployed in the real-world settings. The human can only wear one motion sensor each leg. And the neural network-based controller can identify -- quickly identify the human motion intention by using the portable -- 2 portable motion sensors in total, generate optimal system profiles to reduce energy cost or muscle effort in the real-world settings. And in the whole of the control process, we haven't tuned any control parameters without any intermediate or the human involved intensive testing. The controller can switch to the different subject, we can see, right, after the training, after the controller has already achieved a good convergence by achieving the maximum reward. So the controller can smoothly switch to the different people, different activities without any intermediate estimation of the parameters tuning or any human involved testing. It can be more generalized, more adaptive in the different participants. In collaboration with Dr. Su's research team, we have conducted experiments on 8 participants. We have -- we found that the neural network-based controller can generate more adaptive assistance profiles for different people, as well as different speeds in the walking and running conditions. And in the total control process, we haven't tuned any control parameters. It is an end-to-end control mechanism, which take the side angle, the hip angle and angular velocity as the input, and direct output, the torque assistance. It has already learned how to quickly identify the motion intention and how to autonomously generate optimal assistance profiles. We found that our learned controller in this closed-loop interaction simulation can significantly reduce the energy cost. We compared with the baseline and powered off and powered on. We found that we can achieve a significant human energy cost reductions in the different activities such as walking, running and stair climbing. We also analyzed the internal behavior of the neural network-based control policy in the different activities in the different participants. This neural network-based control policy directly developed many localized clusters of states, which corresponds to the different gait phases in 1 gait cycle to represent the distinct events. The key point in the 1 gait cycle, we analyzed the internal state, including the joint angle, joint angular velocity and also the [indiscernible] learned from the simulation. So, this implies that this neural network already learned how to autonomously assist people in the -- also in the different individuals. We published this work in the Nature in collaboration with Dr. Su's research team. In my future work, my Robotics and Intelligent Learning Lab are working on the personalized neuromechanical simulation for the people with mobility disabilities. We have achieved some interesting results on the neuromechanical simulation for the people with arthritis. Our final goal is to investigate the learning-enabled safe, assured autonomous system in the community settings. We have focused on the additively manufacturing, the neuromechanical simulation, also like including the image processing and also the biomedical imaging processing with application in the healthcare and space exploration. Thank you so much.

Thank you. Thank you very much, Shuzhen, for this very fascinating talk. And we'll move on to our last speaker. So, we welcome Emmanuelle Brès. She is a Research and Development Engineer at Wandercraft, one company commercializing advanced powered exoskeleton. And her presentation today will be titled AI-Powered Exoskeleton Revolutionizing Rehabilitation and Mobility, as the title of our special session. So, Emmanuelle?

Hello, everyone. Thank you for the opportunity to give this presentation. Indeed, I did not look very far for the title, but I'm very happy to be here to tell you all about AI-powered exoskeleton revolutionizing rehabilitation and mobility, and more specifically, what we do here at Wandercraft. So, Wandercraft was founded in 2012 by 3 engineers, Matthieu, Nicolas and Jean-Louis. Nicolas and Jean-Louis' families have rare genetic disease. So it's really a great motivation to have the family of colleagues that are coming in to test the device, give us feedback and orient us. And it's a great mission to be able to focus on the SDGs of reduced inequalities, industry innovation and infrastructure, and good health and well-being. So Wandercraft's vision was always to restore some mobility and some independence to avoid the health issue that come from sitting all day. So, 30 million people are impacted by this because 30 million people are using a wheelchair worldwide. For a small time line, so like I said, 2012, the company was founded. And then, in 2019, we got our first clinical device certified in Europe. We had to wait until 2022 for FDA clearance for cardiovascular accidents. And then, we had the launch of Atalante X, which is our first device that I will detail a bit later. We then got an extension of inclusion to SCI T5 to L5 and upgraded feature to Atalante X, and then opened U.S. headquarters in New York City, where we are currently doing clinical trials for our personal exoskeleton, which is our second device that I will detail later. In 2024, we also had hyper growth in Europe with 80-plus rehab centers that were deployed. And we're hoping in 2025 to have the certification of our personal exo and to further develop -- expand the indications and upgrade our devices already on the field. So, the first device, Atalante X, is an innovative technology that replicates human gaits. I don't need to tell you how important it is for health reasons to be as close to human gait as possible. One aspect we were really focused on and we're I think the most proud of is that it's a self-balancing device, so no crutches are needed. This means that the patient can entirely focus on the rehabilitation of their lower limbs and are not limited by the fatigue of their upper limbs. So it's an anthropomorphic device. It has 12 mechanical joints. So especially, the hip joint has a sagittal, transverse and frontal joint, which is mimicking the ball-type joint femoral neck in the hip. This allows us to have good stability. Especially when turning, we're able to turn 180 degrees, not from the same spot basically. We also have sagittal knee joints and sagittal and frontal ankle joints and 2 prismatic joints for the tibiae and the thighs so that the exoskeleton can be adjusted to the patient's limb since it is used in medical centers where a lot of patients try the exoskeleton. So, here is a small video of a session with Atalante. So this is my colleague, [ Jeanne ], and she is installing a patient inside of the device. So, as you can see, she can -- it's a tool for a therapist, so they can easily make the patient stand up. And then, it's a multi-directional gait exoskeleton. You're able to take steps on the very first session. There's also an active mode, where you can decide the amount of assistance that you use so that you can stimulate the patient, depending on their pathology and their current medical needs. And so, Atalante, we are aiming to revolutionize rehabilitation through our partners in clinical research and robotics lab across the world. We have 100-plus centers globally. There is more than 1,000 patients that have taken steps in Atalante, and 5.6 million steps have been done in 2024 so far. There's still 2 months remaining. And like I mentioned earlier, we are now a 2-technology company because we worked on Atalante for many years, but the vision was always to do a personal exoskeleton, which is similar to Atalante in the meaning that it's self-balancing and there's no need for crutches and it's focused on anthropomorphic gait. But this one is for personal use. So we want to put the accent on mobility outside, independence, social aspects, while still improving the user's health. And so, we are hoping to launch it soon. So, as you all know, the key technology for stability is a very powerful sensing ability. So our exoskeleton has been fitted with 27 sensors that gather information like inertial, angular and force measurements. With an exoskeleton of this size with thin long legs and supporting the user's weight, it's a very complex system, and it twists and flexes under the weight. But thanks to the fusion of all the data of the sensor, we're able to observe and react to this deformation, and we process more than 75,000 data per second. Here is a short video. You can see, we use a dummy for experiment at the office. And we focused on making a device that fuses with the user's body. It also has 12 degrees of freedom, 12 joints and 12 motors, and it is self-balanced, and its architecture allows a large variety of motion like standing, walking, climbing stairs or squatting. We've been lucky enough to show it in our hometown in Paris for the Olympics and Paralympics torch relay. So, this is Kevin. He is a pilot that comes in often to test the exo. He is a full paraplegic, and he is helping us get the feedback on the exo. So, that was a great day that I wanted to share with you. So what is the future work we're hoping to do? As you can imagine, I'm not at liberty to go too much into details. But of course, everybody has seen how AI is extremely powerful, and we are currently testing its potential. So, for example, one of the applications could be through stabilization. We've had a published work on push recovery on the exoskeleton, Atalante, so our first device using reinforcement learning. We are -- since we want to use the device outside, we need to be able to react to many types of grounds and also to, yes, external pushes because the most important thing to us is our users' safety. Another application would be stair climbing. So this is a presentation of a first -- a video of a first concept that we have for climbing stairs, but we still want to improve it and robustify it. So, that was -- that is still Kevin in New York that is climbing up and down stairs. So you can see here, it's a great example of how not having crutches is really useful because then you can grab things, you can really go on with your life and have a lot of independence. And yes, you can see the downstair motion here. And finally, one other aspect where we want to include AI in our future version of exoskeleton is for vision for obstacle avoidance. So we want the use of the exoskeleton to be as worry-free as we can. So we don't think about the fact that there's an obstacle in front of us and we want to avoid it. So we are going to include cameras and use vision to try to adapt the trajectory and the motion to the environment of the user. So I thank you for your attention and give the floor back to you, Olivier.

Fantastic. Thank you so much, Emmanuelle, and thank you to all of you for these very fascinating presentations. And also thank you for keeping up to the timing. So now we have about 45 minutes for -- 40 minutes for opening the discussion part. So I will ask everyone to turn on their video again. And I'll remind the audience that they can ask their question on the interface and that they will be brought back to me to be including into the discussion. I'll be happy to take few questions from the audience, if that's possible. And I would also like to welcome Dr. Katie Strausser, who is a Principal Controls Engineer at Ekso Bionics. It's another company developing powered exoskeleton. So welcome, Katie, for the discussion part. And since you didn't have the opportunity to give a presentation, I'd like to give the word to you in the first place, ask you to comment, if there's anything you would like to add with respect to what has been said so far and maybe link it to the activities you are doing at Ekso Bionics.

Thank you so much. Thanks for allowing me to be a part of this. I'm so impressed with the work that's being done in AI with regard to exoskeletons. And I know that our patients are very excited about it as well so that we can provide better outcomes for them. I'm really excited to hear all the talk about model-based training because I think that's one of the kind of roadblocks to a lot of AI is that most of our patients aren't in the hospital or don't have the time or the ability to train an exoskeleton for good outcomes. It looks like most of the work at this point is being done with more of an able-bodied model, which makes sense. But most of the patients that we've seen some of the best outcomes with rehabilitation exoskeletons are the ones who have severe compensatory strategies, whether they're pushers or they have contractors or they're used to doing odd things with their hips or something to accommodate for their weaknesses. I was just wondering if we could talk about how we can start to incorporate some of those deficiencies into our models to provide better outcomes in rehab.

That's a great topic. I guess we should pass the question to Shuzhen and Hao Su.

Okay. I can explain this first. So I think yes, we try to develop the closed loop the human interaction simulation, right? We can try to like use -- only using the model, mathematic model and the data-driven, the reverse learning. We can generate a high fidelity, I mean the human agent in the simulation. So this agent can generate -- can provide the -- I mean, can provide a realistic response to the robot. So that is the first thing. As long as we can have like accurate mathematic modeling, if that means in this training process, we can generate sufficient training data for the exoskeleton controller. right? We don't need to use -- I mean, the large data site, we don't need to use a large amount of data to turn on your network. Since we already created -- built the accurate mathematic modeling in the simulation. So we can create -- so for example, I can create [ tongue ] human agent in the simulation. So that means I can get a large amount of data. And this tongue human agent can simultaneously like interact with the robot. So I can get like many large amount of data for the training of the neural network-based controller. So -- and yes, I think we also want to extend this framework to the people with like mobility disabilities. And in my current work, we are also focusing the personalized simulation. So that means I can collect one subject data without any device, only like work in the motion capture live, only collect 5 minutes data, including the joint angle, the angle velocity or the ground reaction falls and also we can use the EMG sensors to collect the muscle activations. We can -- I can use the data to achieve the personalized human agent for this subject. So in -- I can still add some the uncertainties in the controller to make sure it can cover one of the real conditions in the real-world settings. It can -- yes, it can improve the likelihood of the -- I mean, of the transfer process since I want to transfer the control policy in the former simulation to the reality. So in this way, maybe we can -- it kind of has much more potential to reduce the human intensive testing as much as possible. We can have like more generalized, more adaptive to this kind of -- to this subject or to this kind of patient with the same symptoms. Yes, thank you.

Thank you for that explanation.

Fantastic. Let me go back to maybe one of the slides that really impressed me and that was in Dr. Su's presentation where there was this time line and then this picture of us being still in the infancy of exoskeleton. So I would like to ask everyone, first of all they agree with this and then what we should do to democratize this technology so that you can really have it everywhere for everyone who needs it, right? So maybe starting with Mohamed.

It's kind of grand question, Olivier, how to -- because...

I'm not here to ask easy questions.

Yes, absolutely. Yes. But the thing is that probably that we now admit that the development of exoskeleton is not that difficult. This is what I would like actually that people are aware about. Regardless of the fact that people are doing -- are developing exoskeletons that are multi degrees of freedom, which is probably not our philosophy. But if we keep it simple in terms of number of degrees of freedom, if we keep it simple in terms of number of actuators, and in terms of components, mechanical components, I think devices as hip knee have this potential to be adopted. The thing is that probably what we see now, the democratization of the devices is related to the price and is related to the adoption from the clinical -- from the therapist which means that we have probably the companies, not only the academia because I'm representing academia, and I have no shares with the start-ups from my group, which makes me very, very independent in terms of opinion. I think that the companies have also to do the efforts for adoption. They have to do the effort in terms of business model, because the business model is likely the barrier, are the main barrier for the adoption because the -- if people -- let's say, people will probably -- will be capable to increase the manufacturing and the production. But at the same -- in the same situation, we will have probably to face another way to sell these devices, which are for now, let's say, around, let's say, yes, $80,000 for the cheapest to something like $150,000 for the one of the -- which is the most expensive. I think the effort is not only from the academia now, but it's also from the other companies. I don't know how -- what Katie and Emmanuelle are thinking about it and about the adoption.

From my point of view, I think that it's really important that we develop the devices the therapist or the user really need, right? So these aren't cheap devices. And I think oftentimes, we as roboticists like to build really cool things and implement really cool learning models, and we sometimes lose sight of the challenges that the therapists are going to face. So what decisions do they need to make, what inputs do they need to give to the device, how do they interact with the device in order to make sure that their patient gets the best outcomes. And then we have to show -- it's on us to show that our device does give the patients good outcomes even if that outcome is just mobility, right? And I think that is an important one. So Emmanuelle give -- showed really great 2 applications, right? We have a device that's used in the therapy clinic, and it's not the same as the device that goes home. They have different needs and stuff. And so often, I think in the early days of exoskeleton development and the use of tech to do this is, we built a really cool thing that was going to do everything. And I think that was back in the day, we had giant phones and they kind of -- of course, I guess it went the other way. They only did one thing and now our phones do everything. So maybe the analogy breaks down there. But -- and I think that's so important for us to think about when we are implementing AI, we can do some really great things, but making sure that we keep an eye on the end user and how they're going to use it and how they're going to implement our algorithms is absolutely critical to your point, Mohamed, so that we can actually sell these devices or else. It's a great academic endeavor, but it doesn't help our patients, so.

Thank you.

Yes. I completely agree with what Mohamed and Katie said. For sure, we need to keep in mind what the end use of the device is going to be. And there's one point I would like to add is that, of course, the device is only as good as it's accessible and affordable. So it's a whole other subject, but this is one of the things where I think the companies and the industry has to work a lot on this is doing research on how useful it is and proving that we should make it accessible and affordable to everybody that needs it.

Yes. If I can chime in, I think it's cost and comfort. I think many exoskeleton is not super, super comfortable and how to make them really comfortable and also reduce the cost. I think still maybe a challenge both for industry and academia, we can maybe solve the problem together by design and also maybe learning another algorithm development. I think I heard a call saying people tend to overestimate the impact technology in the short-term and about underestimate the impact technology in the long-term, right? So if you look back maybe 10 years ago, the robot or exoskeleton really advanced a lot, right? It's still kind of the 10 years ago, what Mohamed showed to us, I think no matter he's the group overall, I think we already made a lot of progress. I think maybe 10 years from now, maybe we can see another big advancement, right? I think maybe we can be a little bit -- we can be optimistic like this. I think also, as Katie said, I think sometimes maybe we consider maybe engineer, developer, we also need to consider user, but also consider the therapists, right? I think in particular for controller, in some way, it's still kind of a very simple called assist and needed control. But I think a therapist have a lot of many other therapy methods, right? I think in Mylan, we are developing some like a curriculum learning, leveraging this -- the Nature paper student and Mylan work together. It's a learning simulation, but we also want to incorporate the curriculum, right? It's not only a very simple, very too simplified assisted and needed control, but it could incorporate many other therapies like therapeutic intervention method, right? So I think learning AI can do a lot of work than only learning simulation, yes.

Okay. So I get a few things out. So first of all, I would need to do another webinar in 10 years to figure out where we stand and then to push forward. So its simplicity, price and accessibility, adoption by the users, different business model, comfort, and I think stick to the needs of the user, right? So and how -- what's the envision and use of the device. So maybe coming a little bit to that. So obviously, this field is highly interdisciplinary, and I guess many of you rely on close collaboration with patients, with clinicians. So how do you foresee the challenges of this collaboration? And -- but also the beauty of it, right, of having patient working with you and seeing the effect, the impact you have with the technologies on these people with disability. Anybody wants to comment on that, both the challenges and the positive -- the rewarding part? Katie?

I'll start. There's no greater feeling than watching somebody who's been sitting in the chair for a long time, stand up and give their loved one a hug. There's nothing better than that. And so every time that we can improve someone's life just a little bit, it's worth a lot of years of research. But every patient is so different. Their diagnosis is so different, how their body reacts to stimuli, how their mind processes things is so different. And we don't I think, fully understand all of the ways that people can react and things like that. And I think that's where -- there's a lot of data out there. There's a lot of ways that we can learn how people are learning. We can learn how their muscles are reacting. We can do a lot of things. And I think that's one place where potentially AI could really, really help us dig through all of that data and figure out how to help our patients better because they are not one thing. They're -- It's so diverse. And so that's a massive challenge.

Emmanuelle?

Yes, I -- sorry. I completely agree with Katie. I think it's very interesting because each patient -- we've had such contradictory feedback, and it's super important not to overfit to one patient. So you've seen it on our video. Kevin is our main pilot. He gives us a lot of feedback. He's great, but we very often have to take a few steps back and realize that he's one person. He has his own pathology. He's not necessarily representative or end user or something, and we need to be able to gather data from more pathologies and not -- and be careful not to overfit to one person. And also, Olivier, you were mentioning the positive aspect of working with patients. And I just wanted to share an anecdote because it's really one of my favorite. We've had a proposal standing up in the exo with somebody saying that usually, people meal when they do this and that for that person, he was so happy to be able to do it standing up. And yes, that was a great example of a great moment for everybody where we felt like we actually had an impact.

Olivier, can I ask a question maybe to Katie?

Okay. Go ahead.

Okay. Katie, according to what you said, do you feel that there is a lot of pressure on the company from the emotional and social point of view that people are, let's say, very happy and very satisfactory with trying with the device? Or do you think that having access to working is still a very nice outcome and is probably enough?

Not sure I understand your question. Sorry, Mohamed.

Okay. The first is that do you feel that there is a lot of pressure on you as a company by having this attempts and this waitings from the patients that they are very happy using the device? Or do you think that having access to verticality to walking as well is still something which is enough?

I mean I do think that we -- as developers of the exoskeletons do have -- that is our responsibility at this point to our patients, right? I do think that -- and we -- it's more than just walking though. We have positioned ourselves as a rehabilitation tool, right? And so it's not just enough to get up and move and to walk. We really have to help to have other outcomes as well. And I think that's -- it's a difficult thing to show. It's a difficult thing to achieve. But I do think it's a really important part of our mission at Ekso Bionics and to make sure that we are improving people's life. It's not just something cool that you get to try, although it is that. And I do think that those moments matter. I think they matter a lot. But somebody walking away from my exoskeleton is much more my goal in many cases. So I'd rather work with them for a while and then bid them adieu. So yes, but it is -- I mean, I do think we have a big responsibility as companies, and I think Emmanuelle would agree that we've taken these products out to market and now it is our -- do they then make them as accessible as possible and as good as possible so that we can help the most patients out there.

Thank you. Emmanuelle, maybe...

Emmanuelle, do you want to respond to that? Yes.

Sorry, I got a little issue with the sound. Do we feel pressure as a company from the expectations of patients? Definitely, definitely. And as important, it is not to focus on one person that is extremely happy and gives us super narrow feedback. We also have to listen to each negative feedback. And sometimes we've put so much effort into it. We have to understand the why. And it is obviously a subject when patients are not able to walk and then we promise them the ability to walk in that device. So sometimes they are disappointed. Sometimes it's not what they imagined. And so we have to deal with that negativity as well. And it can be tough. But in the end, we need to remember the big picture. And of course, we need to always wonder, okay, why that person was disappointed and what can we do to make it better. And it's sometimes tough, but negative feedback is very -- it's almost more important than the positive one because it allows us to improve ourselves.

Yes. I mean going a little bit along those lines also taking into account some of the questions that I received from the audience. So one of the aspects that was mentioned several times, and I know that it's a little big challenge in exoskeleton development, it's the personalization of it was mentioned several times, right? So personalizing it to a solution that fits very well one specific user versus developing something that is available for as many people as possible. So of course, this is a trade-off. This is a big challenge. So how do you try to balance these aspects in the developments you're doing? Maybe I start with Mohamed.

Thank you, Olivier. The thing is that, okay, personalization is something that we still consider as a challenge. You asked also something related to that before, how to use AI actually to be more effective with respect to users. No, Katie, did ask this. And she said, okay, that probably that -- what we did propose is something which is mainly related to healthy users and I think how we would like actually to do it with people with impairment. So I do consider it as still a challenge. We have some approaches for -- with healthy still for now, in which we try to personalize the parameters of the controller. We have some ideas actually how also to -- for partial assistance, not for mobilization. For partial assistance to do the parameterization of the controller with some parameters and then to define with cost functions to try to define the outcomes and to have this iteration to do some optimizations. We have some ideas that we are trying to implement these kind of things. We are also working on functional electrical stimulations. And we are still -- we are also using the same approaches by trying to find the personalization and to do this cartography and this -- this mapping to personalize this mapping between the required outcomes and the -- what happens and then to have this deviation, which helps us to use some optimization functions and to do it. But as I said, it's still something that we are trying to do, which means that I'm not capable to answer that to say, okay, that we are doing it in that way, and it works very well. It's still [indiscernible] from our side.

What about the level of the companies, right? I think usually personalization is challenging, especially in terms of hardware, right? You cannot really have too many adaptation that can be made. So how do you handle that from the point of view of a commercial product? Emmanuelle?

Okay. I'll go. Yes, I was going to say, I think your question worth in 2 aspects, the software one. So the software one is going to be just math and working on very complex algorithms that are robust. But I think it's more maybe interesting now to talk about the hardware aspects of it. And this is, of course, very complicated. To us at Wandercraft, our first focus is always going to be safety. So especially now that we're working on the personal exo that is walking on the street, that is going to work on pavements to be pushed by people. We have this first hardware constraint that is going to be, yes, safety and making sure that nothing can happen to the patient that is the user that is inside of the exo. And so one aspect is that it's like, how you say, seat belt in cars. We need to make sure that the user is going to do and respect every hardware element that we added to make sure that the patient is safe. So in a way, I think like we've already seen some of our users trying to kind of adapt the exo to them like using a hook to put their handbag on or something like this because every patient is different. They have a very different strength in their upper limb muscles, for example. And we are working very hard in making very clear what the instructions are and trying to -- even in a design kind of approach, making sure the device can only be used the way we intend it to be used. And then, of course, we want to personalize it as much as possible to the different pathologies. But I think, to be honest, this will come in a second step. Once the device is out there used and we have some feedback, then we can evaluate if it's worth it to make 2 different versions or something like this. But for now, yes, our main interest, our main point of attention is so that most -- the bigger number of patients can use the device. So we want to be -- we're making the device for the person that has the highest level of pathology and that the patient remains safe in the exo.

Okay. Thank you. Maybe since we only have about 15 minutes left, I want to go back to the artificial intelligence topic and maybe ask you a bit of a very general question on exo. You've been working on this field for some years already. So now with the advancement of artificial intelligence, what do you see -- what has mainly changed in how you approach development of exoskeleton, what is the biggest potential that artificial intelligence is bringing to this field according to you? Maybe we start with Shuzhen.

Yes. I have seen -- in recent years, I have seen many advancements in the machine learning-based control in the robotics field, some advanced algorithm generated from the computer science field, right, like large language model, natural language processing, image classification, all the machine learning algorithms. So yes, we are trying also -- we are all of the experts we are trying to investigate the multidisciplinary field. So I think we can -- in the future, we can continue to focus on using the advanced the algorithms in the -- I mean, the machine learning, in the computing-based technology. You can use the strong power, computing or simulation, maybe we can -- to design the personalized device for a lot of people, not for the -- I mean, the specific subject, even for the industry or the academic field. So yes, I mean, also the main challenge in this field is we -- since the final goal, we want to design the personalized device in the community settings for the people with disabilities. And we -- even though we already have some data site from the healthy -- from able-bodied subject, I think the main challenge in the -- on the people with disabilities is the lack of the sufficient training data from the patients. And based on my experience, I need to collaborate with many professors from other universities from the hospital or health system. I can get some -- the image or the data or they can recruit the patient for us. So that means, I will collaborate with much more -- many professors and many experts like physical therapy and other research experts. And I also need to get much more funding to support this, to support this research, so we can investigate the personalized device for the -- in the future. So I think the main challenge in this field is still the data site, the sufficient training data site for the people with disabilities. Yes. That's my comment.

Okay. Thank you. And Dr. Su?

Yes. Thank you. I think maybe there are 2 major new trends about AI. I think first is the physical AI. I think kind of embodied AI, right, not AI algorithm itself, they need a body, no matter exoskeleton or humanoid, right? I think the first thing is the physical AI. The second is human-centered AI. I think AI is powerful, but we don't want AI to take people's job, right, how AI can really have a human, right? But if we really think about the technical advanced opportunities, I think we talked about personalization, talked about safety. AI is useful, but I think it's still kind of in the early phase for exoskeleton researcher or developer or industry to really embrace AI. But I do see maybe 2 opportunities for AI for our community. First is maybe how can AI can improve personalization. In our Nature paper, we talk about same to real, right, how we can use simulation to have -- to minimize doing a lot of experiments, right, and make things more cost, more affordable and efficient. But there's also real to same, right? You think about real to same, same to real. When we say real to same, I mean, like, for example, we can use a smartphone to take a video, right, or person's gate, people's gate and to generate data. Even their smartphone, there's some app called OpenCap developed by Stanford University. We can generate the data. And even this kind of movement data can actually quantify neurological -- what happened in the brain in some sense, right? It's not everything, but kind of motion is also -- movement can indicate a lot of information. The second point is about the safety, how AI can help with safety. In our simulator, we work with Shuzhen also another faculty from New Jersey Institute of Technology, Alex Zhou. We are working -- developing how to simulate falling some kind of corner cases, some risky cases. And it's very challenging or impossible to ask a real person to do falling and people do fall even without exoskeleton, right? So how we can leverage learning to simulate this kind of challenging situations, right, to really understand the controller, understand the safety. I think that's another maybe promising solution. The third one maybe, as I mentioned, the curriculum learning to do some like a therapist like curriculum to really empower AI with some intelligence like a human, right, not only very simple, oversimplify the controller, but really human therapist control strategies. Yes, thank you.

Yes. Thank you. I'll just ask a question from the audience, which relates specifically to your model. So -- the question is how far do you think we could go with user-agnostic controllers in terms of user anatomies with only relying on the movement signals? Would it still work for a person with extremely different segment inertia? What has your experience been on that? So Dr. Su or Shuzhen, one of the 2.

Okay, I can try to answer this question. So yes, as country, we are like working on the autonomous controller for the able-bodied subject. And the reason we only use the portable motion sensors to detect human intention because they are like low cost, accessible and very easy to deploy in the hardware in the real-world environment. So yes, it is like should be easier to -- I mean, to use the portable motion sensors to detect human intention. But for the -- I mean, for the people with disabilities, right, if we want to focus on the I mean, the body shape or hip joint anemical like the structure, I mean, focus on -- also focus on the physiology, not for the -- not only for the motion sensors. So we are also trying to extract some features by -- the feature by analyzing, I mean, the image, the biomedical image for the people with arthritis. So by analyzing there the joint shape, the structure in the simulation, we can also try to create the personalized simulation for -- in terms of anatomical structure in the simulation, not only focus on the motion sensors. So I think it should be, we need to consider the applications. We also need to consider the population of people. So yes, we can also try to add I mean, more the detailed structure for the human body in the simulation, not only like using portable motion sensors. Yes, that's like -- I try to explain this, yes.

Yes. I think the short answer is for us right now, it's kind of a one-size-fit-all solution. And only right now, we probably focus on testing on able-bodied population. I think it works and it works pretty well, but maybe it's not optimal if you really want to get some optimal control strategy, right? I think the next step is try to move to maybe more personalized control, as I mentioned, like using real to same, and same to real kind of -- this kind of learning methods and data-driven learning method and try to personalize this no matter for able-bodied population, and probably it also works for people with disabilities, right? We don't quite know this. Very likely, it works. As the question said, maybe the movement information already because a lot of human control strategy, right? So probably it's good enough. So we don't know yet. But maybe pretty promising if we continue pursuing this kind of direction, yes.

Yes. Also I'd add one comment in it. So since I create, I mean, controller in the simulation, right, focus on the able-bodied subject. But actually, I also have tested the older adults and also the people with hip OA. So yes, so I think some patient also give me some response, they can also feel the correct -- I mean, the adaptive assistance in the correct timing. So I think we -- as like Dr. Su said, it is not optimal, but it can be more adaptive, it should be useful, not only for the able-bodied, also for the people with some mobility disabilities. Yes, in the future, if we only focus on some specific type of the -- for some disease. So yes, so that means in the simulation no matter we use the same to real, real to same. We need to make sure in the simulation and training should be personalized to the specific type of the patient. Yes, thank you.

Good. Thanks a lot. Maybe as one of the last final questions, I'll go back to maybe many people that are listening to us are academics that work in the field and are interested to maybe once create their company related to these types of technologies. I'm going to ask people who work in the companies to tell us about the main challenges according to them, what are the main hurdles when you want to move from research prototypes to medical product. I know it's not easy. So many challenges. Katie?

So, yes, moving from the research lab to the field is really, I mean, one of the biggest things is device reliability, right? When we're working in the lab, we are okay with zip ties and duct tape. Our customers are not. So making sure that our devices work as expected. And as expected, is a kind of hard thing to figure out, right, because what we expect as engineers is not always what a therapist may expect. And so making sure that we're either training or communicating clearly or understanding their needs really clearly is very important. And then there's -- I think one of the biggest challenges is constantly adapting to what the therapists need while not overfitting to a specific therapist. But we constantly can collect data from our devices and we can analyze sessions for specific individuals and help give feedback to the therapist if something went wrong, if it was a device problem or a use problem or something like that. I do think that's somewhere where AI could be really powerful. There's a lot of data coming out of these devices and can that be used to bridge the gap of work in the lab directly with the engineer who built it versus 800 devices out across the world, I can't work with every therapist. And so is that somewhere where we could use some of these data systems and help our therapists more, but maintaining that usability, that adaptability to all of our patients is so important. And so that's -- it's a huge challenge and very different from moving from having 3 or 4 test pilots when we were in the lab days to 800 devices globally. It's a big change.

Yes. Emmanuelle, do you want to comment on that as well? Main challenges.

Yes. The word I had in mind was definitely reliability, maybe software or hardware. Of course, as developers will know, we need to do tests. We need to think of all the test cases possible, and then we'll forget some. And so that's also what we need to be always careful for. And the hardware reliability as well. We -- at Wandercraft, for example, we use -- we try to do like thousands and hundreds of thousands of steps. And then there's always something that will come up that we hadn't expected, and I think this is very precious. And in the reliability aspect, there's also all the regulations that take place. It's a good thing that there are so many regulations. But if you want to go from a prototype to an actual industrialized and commercializable product, you will need to listen to the regulation, and that takes a lot of time and a lot of effort and a lot of accent on the reliability.

Very nice and we are not even touching the regulatory aspects and also, of course, very challenging when you move from the lab to a company, right?

That's a whole another webinar.

Exactly. Maybe we don't open that box. Okay. Maybe finally, a last question, you -- maybe to Mohamed. So we've had -- I want to just briefly mention the initiative of CYBATHLON that took place in Zurich like a month ago, probably you're all familiar with that. So there was a poor exoskeleton race, right? You participated to it. You mentioned it in your -- in one of your slides. What's your take on these types of initiative and maybe from your own participation, what do you feel the field has gained from -- or you have gained as a participant with these types of initiative?

The thing is that you are pushing someone who is convinced me of the importance of this. Yes, the thing is that, yes, this is -- I believe, let's say it in that way. I believe that these kind of initiatives about public awareness about assisted technology. But with the idea, let's say, to bring the devices, to bring the -- what we are developing, not to the clinical implementations, but to bring them actually to daily living activities. So if one of the motivations is to do that, probably this kind of initiatives of acceptance is the most that we can do. We have probably to find ways how to finance actually the involvement of the labs because we discussed this. And I said, for example, in our case that it costs too much actually to be involved in a race by having a team which is involved in the training with the device, with the patient, that takes a lot of time, which means that we have probably also to find a way how to promote the universities to be involved in that kind of events because as well as I do understand that companies may be interested in this kind of events. But for the universities, it's probably too -- not that easy actually to be involved with this -- according to the financial requirements. But I'm totally agree of the importance of these kind of events.

Anybody else wants to comment on that or would I have to raise another point?

No, I can just say that we participated in the previous CYBATHLON session. And it was amazing. It brought us a lot of feedback. And I do agree that these events are so, so important. And they're also important because they popularize something and then they advertise it and a lot of people -- yes, it -- some people that are not so interested in the subject can become interested because it's also done in a very fun way, I would say.

Absolutely.

Exactly. Yes. Okay. Well, with this, I would like to thank you all for your participation. I think we had a great discussion. There are still some questions open, then I think you are welcome to join on the neural interface to interact with more with the audience if you have the time. But from my side, I would like to close here, and thank you again very, very much for sharing with you -- with us our -- your research, and thanks again for everything. And I hand back to Guillem.

No, thanks truly for all of you. I mean we really enjoy these highly interactive sessions with all the panelists and the audience. So thanks, Olivier, for accepting again this role of moderating and like one of our AI events. And a big thanks to Mohamed, Hao, Shuzhen, Katie and Emma for participating in today's session and to all of the participants to make this such an engaging discussion. And we really encourage you to check our AI for Good program online to see more robotic sessions that may be of interest to you. And this is the end of this section and the start of a really brief networking. If you are still available on the neural network, there are still some pending questions. And if not, see you next time. And I would now give the floor back to Anna for the closing information. Thank you very much. Have a nice evening. Bye-bye.

Thank you for participating in today's AI for Good session. We hope you've learned something new, innovative and engaging in today's event. We now encourage you to continue the conversation on the live video wall in the neural network. Here you can ask questions, like and comment, share links, complete the poll, connect with interesting profiles, or speak one on one using the chat and video function. We invite you to explore the lobby, try the smart matching quiz, visit the virtual exhibits, poster boards, the eShop, and build your personalized AI for Good program. Let's shape the future of AI for Good.