PERSPECTIVE published: 16 December 2021 doi: 10.3389/fnbot.2021.750519 Adaptation Strategies for Personalized Gait Neuroprosthetics Anne D. Koelewijn 1*, Musa Audu 2,3 , Antonio J. del-Ama 4 , Annalisa Colucci 5 , Josep M. Font-Llagunes 6,7 , Antonio Gogeascoechea 8 , Sandra K. Hnat 2,3 , Nathan Makowski 2,9 , Juan C. Moreno 10 , Mark Nandor 2,11 , Roger Quinn 2,11 , Marc Reichenbach 12,13 , Ryan-David Reyes 2,3 , Massimo Sartori 8 , Surjo Soekadar 5 , Ronald J. Triolo 2,3 , Mareike Vermehren 5 , Christian Wenger 14 , Utku S. Yavuz 15 , Dietmar Fey 13 and Philipp Beckerle 16 1 Biomechanical Data Analysis and Creation (BIOMAC) Group, Machine Learning and Data Analytics Lab, Faculty of Engineering, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, 2 Department of Veterans Affairs, Louis Stokes Clevel and Veterans Affairs Medical Center, Advanced Platform Technology Center, Cleveland, OH, United States, 3 Department of Biomedical Engineering, Case Western Reserve University, Cleveland, OH, United States, 4 Applied Mathematics, Materials Science and Technology and Electronic Technology Department, Rey Juan Carlos University, Mostoles, Spain, 5 Clinical Neurotechnology Lab, Neuroscience Research Center (NWFZ), Department of Psychiatry and Neurosciences, Charité - Universitätsmedizin Berlin, Berlin, Germany, 6 Biomechanical Engineering Lab, Department of Mechanical Engineering and Research Centre for Biomedical Engineering, Universitat Politècnica de Catalunya, Barcelona, Spain, 7 Institut de Recerca Sant Joan de Déu, Esplugues de Llobregat, Spain, 8 Department of Biomechanical Engineering, Edited by: Faculty of Engineering Technology, University of Twente, Enschede, Netherlands, 9 Department of Physical Medicine and Zlatko Matjacic, Rehabilitation, MetroHealth Medical Center, Cleveland, OH, United States, 10 Neural Rehabilitation Group, Department of University Rehabilitation Institute Translational Neuroscience, Cajal Institute, CSIC, Madrid, Spain, 11 Department of Mechanical Engineering, Case Western (Slovenia), Slovenia Reserve University, Cleveland, OH, United States, 12 Chair of Computer Engineering, Brandenburg University of Technology Cottbus-Senftenberg, Cottbus, Germany, 13 Chair for Computer Architecture, Department of Computer Science, Faculty of Reviewed by: Engineering, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany, 14 IHP-Leibniz Institut Fuer Innovative Carlos A. Cifuentes, Mikroelektronik, Frankfurt (Oder), Germany, 15 Biomedical Signals and Systems Group, University of Twente, Enschede, Escuela Colombiana de Ingenieria Netherlands, 16 Chair of Autonomous Systems and Mechatronics, Department of Electrical Engineering, Artificial Intelligence Julio Garavito, Colombia in Biomedical Engineering, Faculty of Engineering, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, Germany Dong Hyun Kim, Korea Advanced Institute of Science and Technology, South Korea Personalization of gait neuroprosthetics is paramount to ensure their efficacy for users, *Correspondence: who experience severe limitations in mobility without an assistive device. Our goal Anne D. Koelewijn is to develop assistive devices that collaborate with and are tailored to their users, anne.koelewijn@fau.de while allowing them to use as much of their existing capabilities as possible. Currently, Received: 30 July 2021 personalization of devices is challenging, and technological advances are required to Accepted: 18 November 2021 achieve this goal. Therefore, this paper presents an overview of challenges and research Published: 16 December 2021 directions regarding an interface with the peripheral nervous system, an interface with Citation: Koelewijn AD, Audu M, del-Ama AJ, the central nervous system, and the requirements of interface computing architectures. Colucci A, Font-Llagunes JM, The interface should be modular and adaptable, such that it can provide assistance Gogeascoechea A, Hnat SK, Makowski N, Moreno JC, Nandor M, where it is needed. Novel data processing technology should be developed to allow for Quinn R, Reichenbach M, Reyes R-D, real-time processing while accounting for signal variations in the human. Personalized Sartori M, Soekadar S, Triolo RJ, biomechanical models and simulation techniques should be developed to predict Vermehren M, Wenger C, Yavuz US, Fey D and Beckerle P (2021) assisted walking motions and interactions between the user and the device. Furthermore, Adaptation Strategies for Personalized the advantages of interfacing with both the brain and the spinal cord or the periphery Gait Neuroprosthetics. Front. Neurorobot. 15:750519. should be further explored. Technological advances of interface computing architecture doi: 10.3389/fnbot.2021.750519 should focus on learning on the chip to achieve further personalization. Furthermore, Frontiers in Neurorobotics | www.frontiersin.org 1 December 2021 | Volume 15 | Article 750519 Koelewijn et al. Perspective on Personalized Gait Neuroprosthetics energy consumption should be low to allow for longer use of the neuroprosthesis. In-memory processing combined with resistive random access memory is a promising technology for both. This paper discusses the aforementioned aspects to highlight new directions for future research in gait neuroprosthetics. Keywords: neuroprosthesis, resistive random access memory, neural interface, personalized devices, perspective, embedded artificial intelligence 1. INTRODUCTION Gait neuroprostheses aim to restore function in persons with paralysis caused by various injuries, diseases, or dysfunctions in the central or peripheral nervous system, e.g., a stroke, cerebral palsy (CP) or a spinal cord injury (SCI). Interfaces with the brain, spinal cord or the periphery are appropriate (Figure 1), whereas the type and location of the interface is highly dependent on the user’s abilities and remaining possibilities for voluntary control. Development of gait neuroprostheses should focus on user-friendliness, and aim to maximize speed and safety, while minimizing fall risk. Furthermore, systems should be portable and easy-to-use to achieve their adoption in real- life environments. Despite some individual success using interfaces with the brain (e.g., Ajiboye et al. 2017), spinal cord (e.g., Wagner et al. 2018), and periphery (e.g., Nandor et al. 2021), widespread application of neuroprostheses is still limited. One of the FIGURE 1 | Overview of the different organs involved in gait and their main challenges are differences between and within individuals. communication structure (left). The right side shows examples of prostheses Between individuals, there is significant heterogeneity in the that interface with these organs, specifically the brain (top), spinal cord (bottom target population, which requires personalization of stimulation, left), or periphery (bottom right). stimulus timing and intensity, and the possible extent of additional motorized assistance. With “personalization,” we mean that every patient is provided with an individualized for each user. Figure 2 provides an overview of the different neuroprosthesis. This personalization can be done on the aims and future directions that are described in this paper. hardware and software level. It can be achieved offline through Movement analyses and simulations are required to understand modeling and simulation, or online through adaptation of the pathological gait patterns and to tailor assistance, while the stimulation scheme to the patient’s current abilities. Personalizing design should also be modular and user friendly. Artificial the intervention is currently a time intensive process of trial and intelligence (AI) would allow control algorithms to be adapted error by a human expert. Automation of this process with expert to each individual user, and potentially over the course of systems that can learn the optimal combinations of muscular rehabilitation. Furthermore, AI has the potential to process data, and motor activation for a given individual would increase e.g., electroencephalography and electromyography, accurately accessibility of gait neuroprostheses (Seel et al., 2016). Over in real-time, which is challenging due to the noisiness of these time, due to fatiguing or improved function after rehabilitation, measurements. Adoption of AI in gait neuroprostheses requires individuals might benefit from different stimulation patterns, development of new software and hardware for efficient on-chip which requires stimulation to be adaptive (Del-Ama et al., 2014). learning, which should also focus on low energy consumption to The development of adaptive interfaces and personalized allow for daily use of the device. devices requires an interdisciplinary approach involving In this perspective paper, the organizers, speakers, and authors clinicians, neuroscientists, engineers, and computer scientists. of the mini symposium “Adaptation Strategies for Personalized The effectiveness of gait neuroprostheses can be quantitatively Gait Neuroprosthetics” which was held during the 10th assessed by measuring the gait speed, where desired gait International IEEE/EMBS Conference on Neural Engineering speeds suitable for community use are between 0.8 and 1.2 (NER) in May 2021, summarize and extend their contributions m/s (Robinett and Vondran, 1988; Lapointe et al., 2001), or and discussions to point out directions for future research. by measuring the metabolic energy expenditure while using The workshop comprised talks and discussion from experts the device (Asselin et al., 2015; Evans et al., 2015; Miller et al., working on assistive and rehabilitative neuroprostheses and 2016). Other acceptable metrics could be gait outcomes such as exoskeletons. This perspective covers three topics: interfacing kinematics or symmetry (Hayes et al., 2020). Furthermore, limits with the periphery, interfacing with the central nervous system, for comfortable and tolerable stimulation should be defined and requirements of interface computing architectures. We Frontiers in Neurorobotics | www.frontiersin.org 2 December 2021 | Volume 15 | Article 750519 Koelewijn et al. Perspective on Personalized Gait Neuroprosthetics for device and control customization to maximize walking ability, and improve our understanding of the interaction between human and device for new movement conditions. We advocate to extend computational neuromusculoskeletal models to encompass paralysis-related muscular constraints (Alonso et al., 2012; García-Vallejo et al., 2016) and auxiliary assistive devices (e.g., crutches/walker, orthoses, exoskeletons, etc.) (Febrer-Nafría et al., 2020, 2021) to achieve subject-specific model-based optimization of the device and its control. Then, such simulations might reveal insights into biomechanical effects such as altered recruitment, reduced force production due to atrophy, fatigue effects, or abnormal synergies (Shin et al., 2018). However, creating a personalized neuromusculoskeletal model of a patient is still challenging, since the exact underlying neurological problems of a patient are difficult to extract, and simulations cannot include the variability in muscle responses over time. Experimental validation of simulated clinical outcomes is therefore required (De Groote and Falisse, 2021; Fregly, 2021). Furthermore, even though simulations FIGURE 2 | Overview of the topics of this perspective on gait can provide insight into the required stimulation pattern, the neuroprostheses. We discuss requirements and future directions for interfaces stimulation needs to be adjusted in practice due to the variability with the periphery and the central nervous system (brain and spinal cord), as in muscle response. well as for the computer architecture. To adapt to the individual user, modular and adjustable hardware and control solutions can be implemented on a joint- need basis, and adapted to different anthropometrics, available muscles and their capabilities, as well as the neurological discuss and provide a perspective on current research and conditions (Del-Ama et al., 2014; Makowski et al., 2021; Nandor future directions. et al., 2021). Control in particular requires an adaptive approach that continuously updates muscle stimulation and electric motor 2. INTERFACING WITH THE PERIPHERY actuation, and potentially prioritizes muscle activity through commanding motors in a trajectory-free ballistic paradigm while Interfacing with the peripheral nervous system is a promising monitoring and managing muscle fatigue. To achieve seamless method to address specific functional deficits (Wilson integration of auxiliary assistive devices, control strategies might et al., 2019). Through user-specific stimulation via surface, rely on sensors onboard the human and the machine to measure percutaneous, or implanted electrodes, muscles can perform a user-device interaction (Lancini et al., 2016; Ugurlu et al., 2020) majority of the work for particular tasks. By implementing hybrid and should favor low impedances so the user’s muscles can actuation approaches combining neural stimulation of affected backdrive the device (Foglyano et al., 2016; Beckerle et al., 2017). muscles with robotic assistance, lower extremity function can Wearable kinematic sensing based on inertial measurement units be restored for patients with gait impairments, who can still is expected to provide equivalent performance to marker-based drive the main features of the walking cycle by their healthy or approaches. To this end, recent works propose to integrate preserved muscles. By bringing together neuroprosthetic and inertial measurement units into musculoskeletal modeling wearable robotic approaches, those hybrid concepts combine workflows (Dorschky et al., 2019, 2020; Al Borno et al., 2021; physiologic benefits by activating the other paralyzed and Guidolin et al., 2021). Beyond this, sensory data from the limbs paretic lower limb muscles through neural stimulation, reducing and muscles could be used to improve control and increase the demand on external robotic assistance, and augmenting muscle perfomance by improved fatigue management compared volitional and programmed movements to enhance safety, to current methods based on muscle activity estimation stability, and endurance. (Alibeji et al., 2017; Mohamad et al., 2017). Neuromechanical simulations of such neuromuscular, biomechanical, and mechanical interactions can be used to optimize the performance of users, hardware design (e.g., reduce 3. INTERFACING WITH THE CENTRAL size, weight or energy consumption), and controller design NERVOUS SYSTEM (e.g., seamless motor/muscle actuation and appropriate sensory feedback) (Alonso et al., 2012; Crago et al., 2014; García-Vallejo Neural interfacing technologies connected to the CNS, the spinal et al., 2016; Uchida et al., 2016; Sreenivasa et al., 2017; Michaud cord (Wagner et al., 2018) and the brain (Ajiboye et al., 2017) et al., 2019; Sauder et al., 2019). Dynamic simulations of the have shown promising results, achieving restoration of control system composed of the human and the neuroprosthesis can of extremities. While myographic control can provide more predict the combined human-neuroprosthesis response, allow reliable and versatile assistance, there is increasing evidence that Frontiers in Neurorobotics | www.frontiersin.org 3 December 2021 | Volume 15 | Article 750519 Koelewijn et al. Perspective on Personalized Gait Neuroprosthetics only stimulation at the CNS level can trigger motor recovery indirectly, by measuring the brain’s energy expenditure (Liew (Soekadar et al., 2015; Cervera et al., 2018). Moreover, when et al., 2016; Soekadar et al., 2021). Invasive brain-computer there is no voluntary control of the periphery, an interface interfaces typically exploit the user’s ability to train the electrical to the central nervous system might be the only remaining activity in their brain, which is recorded by electrocorticography approach to restore movement or communication (Birbaumer or multielectrode arrays (Hochberg et al., 2012; Ajiboye et al., et al., 2014). Most neural interfaces focus on either the brain or 2017). The future of individualized brain-computer interfaces the spinal cord, but recently, pioneering studies also investigate interventions relies on advanced algorithms for automated combinations, either of the brain and the spinal cord, or the brain detection of brain states and self-adapting neurofeedback, and the periphery (Dixon et al., 2016; Shulga et al., 2021). as well as on hybrid neural interfaces, integrating different For the spinal cord, a key goal is to interface with specific biosignals, e.g., electroencephalography, electromyography, or motor circuitries responsible for locomotion. Commonly, electrooculography, to allow for a more robust and safe control invasive interfaces are used (Hochberg et al., 2006, 2012; in real-life applications (Witkowski et al., 2014; Soekadar Wagner et al., 2018), which require complex setups to infer et al., 2016). To enable a broad adoption of brain-computer neural information of movement intention. Noninvasive interfaces in real-life environments for clinics and home use, approaches, which combine high-density electromyography, portable and easy-to-use systems need to be designed, requiring blind source separation and neuromechanical modeling, comfortable electroencephalography-headsets that minimize could offer an inexpensive alternative. Although high-density preparation time and allow self-applicability. Furthermore, new electromyography records spatiotemporal myoelectric activity machine learning approaches are needed to optimize calibration at the periphery level, it underlies interfering information from time without inflating the number of sensors. spinal neural cells (i.e., alpha motor neurons). Blind source Instead of focusing on interfacing solely with either the brain, separation (Holobar et al., 2009; Negro et al., 2016) enables central nervous system, or periphery, the next generation of separating the interfering activity from neural sources, thereby gait neuroprostheses for movement rehabilitation may aim to retrieving the activity of actual motor neuron pools. In turn, develop an interface on multiple levels. The advantage is that the decoded neural firing events can be employed to drive the paired activation of pre- and post-synaptic motoneurons comprehensive neuromusculoskeletal models (Sartori et al., at the level of the spinal cord is crucial for facilitating the re- 2016, 2017). Such an interface allows for the activity of actual wiring of functional connections after spinal cord injury (Dixon motor neuron pools to be decoded by separating the interference et al., 2016; Shulga et al., 2021). Motor cortex activity can also activity from the neural source. This technique is currently used be used to control spinal cord stimulation (Capogrosso et al., for characterizing motor neurons during voluntary contractions 2016) instead of external control using a mobile app (Wagner (Farina and Holobar, 2016) and reflex movements (Yavuz et al., 2018). Adapting the parameters of peripheral stimulation et al., 2015) in healthy and impaired (Holobar et al., 2012) to the ongoing neural activity has also proven to play a key individuals. Moreover, other applications include studying the role in fostering neuroplastic reorganization (Mrachacz-Kersting neuromechanical response to external devices (Farina et al., 2014; et al., 2016; Bonizzato et al., 2018). Therefore, we envision Gogeascoechea et al., 2020). This technique can therefore extend the implementation of brain-computer interfaces that allow for current open-loop rehabilitation techniques into closed-loop personalized adaptive modulation of brain activity and alpha neuro-modulative approaches. However, its use is mostly limited motor circuitries. to isometric contractions or slow dynamic contractions, mainly due to computational challenges related to the assumption that motor units are stationary and real-time implementation 4. INTERFACE WITH THE EMBEDDED of the method. Both model-free AI (e.g., machine and deep COMPUTER ARCHITECTURE learning techniques) (Chen et al., 2020; Clarke et al., 2020) and model-based techniques (e.g., data-driven mechanistic The main requirements for interface computing architectures modeling) (Sartori and Sawicki, 2021) are explored to enable of gait neuroprosthetics are that the system is personalizable real-time implementation, which would allow mechanical and and adaptable, it should consume as little energy as possible, neural adaptations to exoskeleton training and neurostimulation such that the system can be used optimally for at least a to be predicted. full day, and data processing should happen in real-time. Besides providing intuitive and seamless assistive control, an Embedded AI for neuroprosthetics is a promising approach important goal at the level of the brain is to promote neuroplastic to achieve these requirements, as it has the potential to be changes and foster functional connectivity between central real-time, while the computing technology and electronics are motoneurons and inactive and/or silent peripheral motoneurons attached to the body or even implanted inside the body. (Donati et al., 2016). By decoding movement intention and gait The importance of a corresponding computer architecture in characteristics in real time, invasive and non-invasive brain- neuroprosthesis has been emphasized (Vassanelli and Mahmud, computer interfaces can directly infer the user’s intention to 2016; Ielmini and Wong, 2018), specifically to develop a move, optimizing rehabilitation outcomes (Soekadar et al., 2015; suitable neuronal architecture to interface with the brain. Mrachacz-Kersting et al., 2016). Non-invasive brain-computer However, no implementations exist so far, and therefore the interfaces can assess large-scale brain oscillatory activity directly, further development and prototyping of concepts is of utmost through electroencephalography or magnetoencephalography, or importance (Mikhaylov et al., 2020). To achieve personalization Frontiers in Neurorobotics | www.frontiersin.org 4 December 2021 | Volume 15 | Article 750519 Koelewijn et al. Perspective on Personalized Gait Neuroprosthetics and adaptability, a main challenge is to perform online learning trade-off between energy consumption for read out and energy on the embedded AI. In contrast, today’s embedded AI only saving from in-memory processing should be further investigated allow inference, while networks are trained offline. This training to minimize overall energy consumption. Furthermore, storage is computationally intensive and consumes much energy. To elements using ReRAM technology have a larger device-to-device achieve low energy consumption and real-time processing for variability than traditional storage elements, which use volatile training as well as inference on the chip, the embedded AI should technology. In particular, this device-to-device variability is a be implemented as low-energy circuits. problem for multi-bit storing (Fritscher et al., 2021b), which An important step to achieve low energy consumption in is desired for embedded AI. Also the variability of switching embedded AI is to use in-memory processing, i.e., operations parameters and energy overheads of analog-to-digital and digital- on data are performed analogously or digitally directly in to-analog conversion is still a challenge for the reliable use the memory. Low-energy consumption is inherent for analog of ReRAMs as well as their comparatively low endurance, i.e., processing, since no digital conversion is required, while the number of allowable switching cycles of the ReRAM cells. specialized hardware architectures without a central processing This is particularly challgengin for neuroprotheses, which have unit can greatly decrease energy consumption since long data higher time series data analysis requirements than e.g., a ReRAM paths are avoided. Using non-volatile storage ensures that no based analysator for detecting atrial fibrillation. To achieve information is lost when power is switched off, what can this, new research should address special training methods that be done when no data processing is required, e.g., when a account for tolerable fluctuations by using learning techniques person wearing a neuroprosthetics stops and does not walk. such as noisy training or dropout layers (Fritscher et al., Accordingly, reloading weights after activation is not required. 2021a). Many international groups currently study applications of non- volatile memory for neural networks and enhanced computing technologies, e.g., utilizing such memory as synaptic elements for 5. SUMMARY AND CONCLUSION artificial neural networks (Hu et al., 2018; Zhang et al., 2020). Resistive random access memory (ReRAM) has compared Neuroprostheses can potentially restore function through to other non-volatile memory technologies, like e.g., Phase external activation of the central or peripheral nervous system. Change Memories, the advantage (Ielmini and Wong, 2018) We have presented our perspective on current challenges of to further decrease energy consumption, due to a unique and future directions in the development of neuroprostheses combination of multi-level programming and a high density in stimulation of the peripheral and central nervous systems, of integration (Milo et al., 2019). Furthermore, ReRAMs and outlined technical approaches to appropriate computer are potentially compatible with the most commonly used architectures, as also summarized in Figure 2. A paramount chip technology, complementary metal–oxide–semiconductor challenge is to restore mobility by effectively combining process manufacturing. Multi-level programming means that neuroprosthetic and wearable robotic approaches and to align multiple bits can be stored on one ReRAM device. Therefore, neuroprostheses to the individual user’s needs and capabilities. weights and matrix multiplications of a neural network can To this end, predictive simulations of the human-machine be stored inside one ReRAM cell, and a neural network can system in dynamic tasks appear to be a promising approach be implemented by saving all network weights directly on the to customize design and control. Still, predictive simulations chip (Perez et al., 2020). As a result, external memory access can only provide rough representations of a paralysis and its is avoided, which can lower the energy consumption drastically highly individual constraints, which cannot yet be covered by (Knödtel et al., 2020). The high-density integration on the chip recent neuromuscular modeling approaches, which, in turn, allows for the ReRAM cells to be closely attached to digital hampers user-specific control and stimulation. We suggest computation units, which decreases energy consumption by that future work should focus on improving neuromechanical avoiding long data paths on the chip. The use of ReRAM simulations of user-device interaction based on experimental technology yielded a 95% reduction in energy consumption for data, e.g., obtaining muscle parameters on the bench and the classification of bio-signals for atrial fibrillation compared in real applications, and aim at prioritizing muscle activity to a traditional approach (Pechmann et al., 2021). Therefore, over robotic assistance. Combining a deeper understanding ReRAM technology is a promising candidate for realization of of neuromechanical dynamics, particularly muscle-group personalized gait neuroprostheses relying on artificial neural excitation through the central nervous system, with multimodal networks in digital (e.g., as ReRAM storage for weights) and sensor networks distributed across the human and the device analog in-memory processing (e.g., as an in-memory processing could foster model-based monitoring and management of element itself). muscle fatigue. However, several challenges exist for ReRAM technology For the central nervous system, real-time data processing, to be implemented on a neuroprosthesis. Currently, storage ease of use of systems, and combining interfacing at multiple of up to three bits on each ReRAM cell is possible (Milo levels are important future directions. Real-time data et al., 2021). However, energy consumption for reading out processing is required to extract useful information from information increases when more bits are stored on a cell. This noisy measurements of brain or spinal cord activity in a useful Frontiers in Neurorobotics | www.frontiersin.org 5 December 2021 | Volume 15 | Article 750519 Koelewijn et al. Perspective on Personalized Gait Neuroprosthetics way. AI is a promising approach to enable this real-time AUTHOR CONTRIBUTIONS data processing, while also allowing for personalization, and limiting the number of sensors required, improving ease of AK coordinated the writing process, the development of the use. Portable and easy-to-use systems should be designed paper structure, and integration of individual contributions. PB to allow for the adoption of neuroprostheses in real-life. and DF supported this for the whole paper. MA, AJd-A, JMF-L, Furthermore, the combination of stimulation at the brain and SH, NM, JCM, MN, RQ, R-DR, RJT, and PB mainly contributed the spinal cord or the periphery should be further explored, to section 2. AC, AG, MS, SS, MV, USY, and AK mainly since these combinations were shown to have benefits over contributed to section 3. MR, CW, and DF mainly contributed stimulation at only one of these levels, and thereby improve to section 4. All authors contributed to the article and approved rehabilitation outcomes. the submitted version. Regarding technological advances of the interface to the computing architecture, a main current challenge is to FUNDING achieve low-energy electronics, real-time data processing, and learning on the chip. Embedded AI has the potential AK is funded by a faculty endowment by adidas AG. to process data in real-time, and allows for learning and MA, SKH, NM, MN, RJQ, R-DR, RJT are supported by inference on the chip. To achieve low-energy consumption,the NSF CPS grant 1739800, VA Merit Reviews A2275-R and embedded AI should use in-memory processing combined 3056, and the NIH (5T32EB004314-15, R01 NS040547-13). with ReRAM technology. To allow the use of ReRAM MS and AG are funded by the European Research Council technology in gait neuroprostheses, we need further research (ERC) under the European Union’s Horizon 2020 research in combining ReRAM technology as multi-bit storage cells and innovation programme (Grant agreement No. 803035). with complementary metal–oxide–semiconductor process AJd-A, JMF-L, and JCM are supported by coordinated grants manufacturing, the commonly-used chip technology. RTI2018-097290-B-C31/C32/C33 (TAILOR project) funded by Furthermore, the functionality of the ReRAM cells should MCIN/AEI/10.13039/501100011033 and by “ERDF A way of be expanded to improve reliability and mitigate the effect its making Europe”. MR is funded by the Lo3-ML project by the device-to-device variability. Federal Ministry for Education, Science and Technology (BMBF) In conclusion, a close cooperation between computer (Funding No. 16ES1142K). AC, SS, and MV were supported by architects, electrical engineers, material scientist, medical the European Research Council (ERC) under the project NGBMI experts, and biomechanical experts is required to (759370), the Einstein Stiftung Berlin, the ERA-NET NEURON design appropriate neuroprostheses that are tailored project HYBRIDMIND (BMBF, 01GP2121A and -B) and the to the user’s need, adaptable, easy-to-use, and consume BMBF project NEO (13GW0483C). little energy. ACKNOWLEDGMENTS DATA AVAILABILITY STATEMENT The authors thank the organizers of IEEE/EMBS NER conference The original contributions presented in the study are included in 2021 for the opportunity to hold the mini symposium in the article/supplementary material, further inquiries can be Adaptation Strategies for Personalized Gait Neuroprosthetics as directed to the corresponding author. part of the conference. REFERENCES exoskeleton-assisted walking in persons with paraplegia. J. Rehabil. Res. Dev. 52, 147. doi: 10.1682/JRRD.2014.02.0060 Ajiboye, A. B., Willett, F. R., Young, D. R., Memberg, W. D., Murphy, Beckerle, P., Salvietti, G., Unal, R., Prattichizzo, D., Rossi, S., Castellini, B. A., Miller, J. P., et al. (2017). Restoration of reaching and grasping C., et al. (2017). A human-robot interaction perspective on assistive and movements through brain-controlled muscle stimulation in a person with rehabilitation robotics. Front. Neurorobot. 11:24. doi: 10.3389/fnbot.2017. tetraplegia: a proof-of-concept demonstration. Lancet 389, 1821–1830. 00024 doi: 10.1016/S0140-6736(17)30601-3 Birbaumer, N., Gallegos-Ayala, G., Wildgruber, M., Silvoni, S., and Soekadar, S. R. Al Borno, M., O’Day, J., Ibarra, V., Dunne, J., Seth, A., Habib, A., et al. (2014). Direct brain control and communication in paralysis. Brain Topogr. 27, (2021). Opensense: An open-source toolbox for inertial-measurement-unit- 4–11. doi: 10.1007/s10548-013-0282-1 based measurement of lower extremity kinematics over long durations. bioRxiv. Bonizzato, M., Pidpruzhnykova, G., DiGiovanna, J., Shkorbatova, P., Pavlova, doi: 10.1101/2021.07.01.450788 N., Micera, S., et al. (2018). Brain-controlled modulation of spinal circuits Alibeji, N., Kirsch, N., and Sharma, N. (2017). An adaptive low-dimensional improves recovery from spinal cord injury. Nat. Commun. 9, 1–14. control to compensate for actuator redundancy and fes-induced muscle doi: 10.1038/s41467-018-05282-6 fatigue in a hybrid neuroprosthesis. Control Eng. Pract. 59, 204–219. Capogrosso, M., Milekovic, T., Borton, D., Wagner, F., Moraud, E. M., Mignardot, doi: 10.1016/j.conengprac.2016.07.015 J.-B., et al. (2016). A brain-spine interface alleviating gait deficits after spinal Alonso, J., Romero, F., Pàmies-Vilà, R., Lugrís, U., and Font-Llagunes, J. M. cord injury in primates. Nature 539, 284–288. doi: 10.1038/nature20118 (2012). A simple approach to estimate muscle forces and orthosis actuation in Cervera, M. A., Soekadar, S. R., Ushiba, J., Millán, J., d., R., Liu, M., et al. powered assisted walking of spinal cord-injured subjects. Multibody Syst. Dyn. (2018). Brain-computer interfaces for post-stroke motor rehabilitation: a meta- 28, 109–124. doi: 10.1007/s11044-011-9284-5 analysis. Ann. Clin. Transl. Neurol. 5, 651–663. doi: 10.1002/acn3.544 Asselin, P., Knezevic, S., Kornfeld, S., Cirnigliaro, C., Agranova-Breyter, I., Chen, C., Yu, Y., Sheng, X., and Zhu, X. (2020). Analysis of motor unit activities and Bauman, W. A. (2015). Heart rate and oxygen demand of powered during multiple motor tasks by real-time emg decomposition: perspective Frontiers in Neurorobotics | www.frontiersin.org 6 December 2021 | Volume 15 | Article 750519 Koelewijn et al. Perspective on Personalized Gait Neuroprosthetics for myoelectric control. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 2020, García-Vallejo, D., Font-Llagunes, J. M., and Schiehlen, W. (2016). Dynamical 4791–4794. doi: 10.1109/EMBC44109.2020.9176362 analysis and design of active orthoses for spinal cord injured subjects Clarke, A. K., Atashzar, S. F., Del Vecchio, A., Barsakcioglu, D., Muceli, by aesthetic and energetic optimization. Nonlinear. Dyn. 84, 559–581. S., Bentley, P., et al. (2020). Deep learning for robust decomposition of doi: 10.1007/s11071-015-2507-1 high-density surface emg signals. IEEE Trans. Biomed. Eng. 68, 526–534. Gogeascoechea, A., Kuck, A., Van Asseldonk, E., Negro, F., Buitenweg, J. R., Yavuz, doi: 10.1109/TBME.2020.3006508 U. S., et al. (2020). Interfacing with alpha motor neurons in spinal cord injury Crago, P. E., Makowski, N. S., and Cole, N. M. (2014). Contributions to muscle patients receiving trans-spinal electrical stimulation. Front. Neurol. 11:493. force and emg by combined neural excitation and electrical stimulation. J doi: 10.3389/fneur.2020.00493 Neural Eng. 11, 056022. doi: 10.1088/1741-2560/11/5/056022 Guidolin, M., Petrea, R. A. B., Roberto, O., Reggiani, M., Menegatti, E., De Groote, F., and Falisse, A. (2021). Perspective on musculoskeletal modelling and Tagliapietra, L. (2021). “On the accuracy of imus for human motion and predictive simulations of human movement to assess the neuromechanics tracking: a comparative evaluation,” in 2021 IEEE International Conference on of gait. Proc. R. Soc. B 288, 20202432. doi: 10.1098/rspb.2020.2432 Mechatronics (ICM) (Kashiwa: IEEE), 1–6. Del-Ama, A. J., Gil-Agudo, Á., Pons, J. L., and Moreno, J. C. (2014). Hayes, S. C., White, M., White, H. S. F., and Vanicek, N. (2020). A biomechanical Hybrid gait training with an overground robot for people with comparison of powered robotic exoskeleton gait with normal and slow walking: incomplete spinal cord injury: a pilot study. Front. Hum. Neurosci. 8:298. an investigation with able-bodied individuals. Clin. Biomech. 80:105133. doi: 10.3389/fnhum.2014.00298 doi: 10.1016/j.clinbiomech.2020.105133 Dixon, L., Ibrahim, M. M., Santora, D., and Knikou, M. (2016). Paired Hochberg, L. R., Bacher, D., Jarosiewicz, B., Masse, N. Y., Simeral, J. D., Vogel, associative transspinal and transcortical stimulation produces plasticity in J., et al. (2012). Reach and grasp by people with tetraplegia using a neurally human cortical and spinal neuronal circuits. J. Neurophysiol. 116, 904–916. controlled robotic arm. Nature 485, 372–375. doi: 10.1038/nature11076 doi: 10.1152/jn.00259.2016 Hochberg, L. R., Serruya, M. D., Friehs, G. M., Mukand, J. A., Saleh, M., Caplan, A. Donati, A. R., Shokur, S., Morya, E., Campos, D. S., Moioli, R. C., Gitti, C. M., et al. H., et al. (2006). Neuronal ensemble control of prosthetic devices by a human (2016). Long-term training with a brain-machine interface-based gait protocol with tetraplegia. Nature 442, 164–171. doi: 10.1038/nature04970 induces partial neurological recovery in paraplegic patients. Sci. Rep. 6, 1–16. Holobar, A., Farina, D., Gazzoni, M., Merletti, R., and Zazula, D. (2009). Estimating doi: 10.1038/srep30383 motor unit discharge patterns from high-density surface electromyogram. Clin. Dorschky, E., Nitschke, M., Martindale, C. F., van den Bogert, A. J., Koelewijn, A. Neurophysiol. 120, 551–562. doi: 10.1016/j.clinph.2008.10.160 D., and Eskofier, B. M. (2020). Cnn-based estimation of sagittal plane walking Holobar, A., Glaser, V., Gallego, J., Dideriksen, J. L., and Farina, D. (2012). Non- and running biomechanics from measured and simulated inertial sensor data. invasive characterization of motor unit behaviour in pathological tremor. J. Front. Bioeng. Biotechnol. 8:604. doi: 10.3389/fbioe.2020.00604 Neural Eng. 9, 056011. doi: 10.1088/1741-2560/9/5/056011 Dorschky, E., Nitschke, M., Seifer, A.-K., van den Bogert, A. J., and Eskofier, Hu, M., Graves, C. E., Li, C., Li, Y., Ge, N., Montgomery, E., et al. (2018). B. M. (2019). Estimation of gait kinematics and kinetics from inertial sensor Memristor-based analog computation and neural network classification with data using optimal control of musculoskeletal models. J. Biomech. 95:109278. a dot product engine. Adv. Mater. 30, 1705914. doi: 10.1002/adma.201705914 doi: 10.1016/j.jbiomech.2019.07.022 Ielmini, D., and Wong, H.-S. P. (2018). In-memory computing with resistive Evans, N., Hartigan, C., Kandilakis, C., Pharo, E., and Clesson, I. (2015). Acute switching devices. Nat. Electron. 1, 333–343. doi: 10.1038/s41928-018-0092-2 cardiorespiratory and metabolic responses during exoskeleton-assisted walking Knödtel, J., Fritscher, M., Reiser, D., Fey, D., Breiling, M., and Reichenbach, M. overground among persons with chronic spinal cord injury. Top. Spinal. Cord. (2020). “A model-to-circuit compiler for evaluation of dnn accelerators based Inj. Rehabil. 21, 122–132. doi: 10.1310/sci2102-122 on systolic arrays and multibit emerging memories,” in 2020 9th International Farina, D., and Holobar, A. (2016). Characterization of human motor Conference on Modern Circuits and Systems Technologies (MOCAST) (Bremen: units from surface emg decomposition. Proc. IEEE 104, 353–373. IEEE), 1–6. doi: 10.1109/JPROC.2015.2498665 Lancini, M., Serpelloni, M., Pasinetti, S., and Guanziroli, E. (2016). Healthcare Farina, D., Jiang, N., Rehbaum, H., Holobar, A., Graimann, B., Dietl, H., sensor system exploiting instrumented crutches for force measurement et al. (2014). The extraction of neural information from the surface during assisted gait of exoskeleton users. IEEE Sens. J. 16, 8228–8237. emg for the control of upper-limb prostheses: emerging avenues doi: 10.1109/JSEN.2016.2579738 and challenges. IEEE Trans. Neural Syst. Rehabil. Eng. 22, 797–809. Lapointe, R., Lajoie, Y., Serresse, O., and Barbeau, H. (2001). Functional doi: 10.1109/TNSRE.2014.2305111 community ambulation requirements in incomplete spinal cord injured Febrer-Nafría, M., Pallarès-López, R., Fregly, B. J., and Font-Llagunes, subjects. Spinal Cord. 39, 327–335. doi: 10.1038/sj.sc.3101167 J. M. (2020). Comparison of different optimal control formulations Liew, S.-L., Rana, M., Cornelsen, S., Fortunato de Barros Filho, M., Birbaumer, N., for generating dynamically consistent crutch walking simulations Sitaram, R., et al. (2016). Improving motor corticothalamic communication using a torque-driven model. Mech. Mach. Theory 154:104031. after stroke using real-time fmri connectivity-based neurofeedback. doi: 10.1016/j.mechmachtheory.2020.104031 Neurorehabil. Neural Repair. 30, 671–675. doi: 10.1177/15459683156 Febrer-Nafría, M., Pallarès-López, R., Fregly, B. J., and Font-Llagunes, J. M. (2021). 19699 Prediction of three-dimensional crutch walking patterns using a torque-driven Makowski, N. S., Campean, A., Lambrecht, J. M., Buckett, J. R., Coburn, J. D., model. Multibody Syst. Dyn. 51, 1–19. doi: 10.1007/s11044-020-09751-z Hart, R. L., et al. (2021). Design and testing of stimulation and myoelectric Foglyano, K. M., Schnellenberger, J. R., Kobetic, R., Lombardo, L., Pinault, recording modules in an implanted distributed neuroprosthetic system. IEEE G., Selkirk, S., et al. (2016). Accelerometer-based step initiation control Trans. Biomed. Circ. Syst. 15, 281–293. doi: 10.1109/TBCAS.2021.3066838 for gait-assist neuroprostheses. J. Rehabil. Res. Dev. 53, 919–932. Michaud, F., Mouzo, F., Lugrís, U., and Cuadrado, J. (2019). Energy expenditure doi: 10.1682/JRRD.2015.09.0188 estimation during crutch-orthosis-assisted gait of a spinal-cord-injured subject. Fregly, B. J. (2021). A conceptual blueprint for making neuromusculoskeletal Front. Neurorobot. 13:55. doi: 10.3389/fnbot.2019.00055 models clinically useful. Appl. Sci. 11:2037. doi: 10.3390/app11052037 Mikhaylov, A., Pimashkin, A., Pigareva, Y., Gerasimova, S., Gryaznov, E., Fritscher, M., Knödtel, J., Mallah, M., Pechmann, S., Perez-Bosch Quesada, E., Shchanikov, S., et al. (2020). Neurohybrid memristive cmos-integrated Rizzi, T., et al. (2021a). “Mitigating the effects of RRAM process variation on the systems for biosensors and neuroprosthetics. Front. Neurosci. 14:358. accuracy of artifical neural networks,” in Proc. 21st International Conference on doi: 10.3389/fnins.2020.00358 Embedded Computer Systems: Architectures, Modeling and Simulation (SAMOS Miller, L. E., Zimmermann, A. K., and Herbert, W. G. (2016). Clinical effectiveness 2021) (Pythagorion). and safety of powered exoskeleton-assisted walking in patients with spinal Fritscher, M., Knödtel, J., Reiser, D., Mallah, M., Pechmann, S., Fey, D., et al. cord injury: systematic review with meta-analysis. Med. Devices (Auckl) 9:455. (2021b). “Simulating large neural networks embedding mlc rram as weight doi: 10.2147/MDER.S103102 storage considering device variations,” in 2021 IEEE 12th Latin America Milo, V., Glukhov, A., Pérez, E., Zambelli, C., Lepri, N., Mahadevaiah, M. K., Symposium on Circuits and System (LASCAS), (Arequipa: IEEE), 1–4. et al. (2021). Accurate program/verify schemes of resistive switching memory Frontiers in Neurorobotics | www.frontiersin.org 7 December 2021 | Volume 15 | Article 750519 Koelewijn et al. Perspective on Personalized Gait Neuroprosthetics (rram) for in-memory neural network circuits. IEEE Trans. Electron. Devices. Soekadar, S. R., Birbaumer, N., Slutzky, M. W., and Cohen, L. G. (2015). 68, 3832–3837. doi: 10.1109/TED.2021.3089995 Brain-machine interfaces in neurorehabilitation of stroke. Neurobiol. Dis. 83, Milo, V., Zambelli, C., Olivo, P., Pérez, E., K., Mahadevaiah, M., et al. (2019). 172–179. doi: 10.1016/j.nbd.2014.11.025 Multilevel hfo2-based rram devices for low-power neuromorphic networks. Soekadar, S. R., Kohl, S. H., Mihara, M., and von Lühmann, A. (2021). Optical brain APL Mater. 7, 081120. doi: 10.1063/1.5108650 imaging and its application to neurofeedback. Neuroimage Clin. 30:102577. Mohamad, N. Z., Hamzaid, N. A., Davis, G. M., Abdul Wahab, A. K., and doi: 10.1016/j.nicl.2021.102577 Hasnan, N. (2017). Mechanomyography and torque during fes-evoked muscle Sreenivasa, M., Millard, M., Felis, M., Mombaur, K., and Wolf, S. I. (2017). Optimal contractions to fatigue in individuals with spinal cord injury. Sensors 17:1627. control based stiffness identification of an ankle-foot orthosis using a predictive doi: 10.3390/s17071627 walking model. Front. Comput. Neurosci. 11:23. doi: 10.3389/fncom.2017. Mrachacz-Kersting, N., Jiang, N., Stevenson, A. J. T., Niazi, I. K., Kostic, V., 00023 Pavlovic, A., et al. (2016). Efficient neuroplasticity induction in chronic stroke Uchida, T. K., Seth, A., Pouya, S., Dembia, C. L., Hicks, J. L., and Delp, S. L. (2016). patients by an associative brain-computer interface. J. Neurophysiol. 115, Simulating ideal assistive devices to reduce the metabolic cost of running. PLoS 1410–1421. doi: 10.1152/jn.00918.2015 ONE 11:e0163417. doi: 10.1371/journal.pone.0163417 Nandor, M., Kobetic, R., Audu, M., Triolo, R., and Quinn, R. (2021). A muscle- Ugurlu, B., Oshima, H., Sariyildiz, E., Narikiyo, T., and Babic, J. (2020). first, electromechanical hybrid gait restoration system in people with spinal Active compliance control reduces upper body effort in exoskeleton- cord injury. Front. Rob. AI 8:645588. doi: 10.3389/frobt.2021.645588 supported walking. IEEE Trans. Hum. Mach. Syst. 50, 144–153. Negro, F., Muceli, S., Castronovo, A. M., Holobar, A., and Farina, D. doi: 10.1109/THMS.2019.2961969 (2016). Multi-channel intramuscular and surface emg decomposition Vassanelli, S., and Mahmud, M. (2016). Trends and challenges in by convolutive blind source separation. J Neural Eng. 13, 026027. neuroengineering: toward “intelligent” neuroprostheses through brain- doi: 10.1088/1741-2560/13/2/026027 “brain inspired systems” communication. Front. Neurosci. 10:438. Pechmann, S., Mai, T., Potschka, J., Reiser, D., Reichel, P., Breiling, M., et doi: 10.3389/fnins.2016.00438 al. (2021). A low-power RRAM memory block for embedded, multi-level Wagner, F. B., Mignardot, J.-B., Le Goff-Mignardot, C. G., Demesmaeker, weight and bias storage in artificial neural networks. Micromachines 12:1277. R., Komi, S., Capogrosso, M., et al. (2018). Targeted neurotechnology doi: 10.3390/mi12111277 restores walking in humans with spinal cord injury. Nature 563, 65–71. Perez, E., González Ossorio, Ó., Duenas, S., Castan, H., Garcia, H., and Wenger, doi: 10.1038/s41586-018-0649-2 C. (2020). Programming pulse width assessment for reliable and low-energy Wilson, R. D., Bryden, A. M., Kilgore, K. L., Makowski, N., Bourbeau, D., Kowalski, endurance performance in al: Hfo2-based rram arrays. Electronics 9, 864. K. E., et al. (2019). Neuromodulation for functional electrical stimulation. Phys. doi: 10.3390/electronics9050864 Med. Rehabil. Clin. 30, 301–318. doi: 10.1016/j.pmr.2018.12.011 Robinett, C. S., and Vondran, M. A. (1988). Functional ambulation velocity and Witkowski, M., Cortese, M., Cempini, M., Mellinger, J., Vitiello, N., and Soekadar, distance requirements in rural and urban communities: a clinical report. Phys. S. R. (2014). Enhancing brain-machine interface (bmi) control of a hand Ther. 68, 1371–1373. doi: 10.1093/ptj/68.9.1371 exoskeleton using electrooculography (eog). J. Neuroeng. Rehabil. 11, 1–6. Sartori, M., Lloyd, D. G., and Farina, D. (2016). Neural data-driven musculoskeletal doi: 10.1186/1743-0003-11-165 modeling for personalized neurorehabilitation technologies. IEEE Trans. Yavuz, U., ,̧ S., Negro, F., Sebik, O., Holobar, A., Frömmel, C., et al. Biomed. Eng. 63, 879–893. doi: 10.1109/TBME.2016.2538296 (2015). Estimating reflex responses in large populations of motor units by Sartori, M., and Sawicki, G. S. (2021). Closing the loop between decomposition of the high-density surface electromyogram. J. Physiol. 593, wearable technology and human biology: A new paradigm for steering 4305–4318. doi: 10.1113/JP270635 neuromuscular form and function. Progr. Biomed. Eng. 3, 023001. Zhang, W., Gao, B., Tang, J., Yao, P., Yu, S., Chang, M.-F., et al. doi: 10.1088/2516-1091/abe3e0 (2020). Neuro-inspired computing chips. Nat. Electron. 3, 371–382. Sartori, M., Yavuz, U., ,̧ S., and Farina, D. (2017). In vivo neuromechanics: decoding doi: 10.1038/s41928-020-0435-7 causal motor neuron behavior with resulting musculoskeletal function. Sci. Rep. 7, 1–14. doi: 10.1038/s41598-017-13766-6 Conflict of Interest: The authors declare that the research was conducted in the Sauder, N. R., Meyer, A. J., Allen, J. L., Ting, L. H., Kesar, T. M., and Fregly, B. J. absence of any commercial or financial relationships that could be construed as a (2019). Computational design of fastfes treatment to improve propulsive force potential conflict of interest. symmetry during post-stroke gait: a feasibility study. Front. Neurorobot. 13:80. doi: 10.3389/fnbot.2019.00080 Publisher’s Note: All claims expressed in this article are solely those of the authors Seel, T., Werner, C., Raisch, J., and Schauer, T. (2016). Iterative learning control and do not necessarily represent those of their affiliated organizations, or those of of a drop foot neuroprosthesis—generating physiological foot motion in the publisher, the editors and the reviewers. Any product that may be evaluated in paretic gait by automatic feedback control. Control Eng. Pract. 48:87–97. this article, or claim that may be made by its manufacturer, is not guaranteed or doi: 10.1016/j.conengprac.2015.11.007 Shin, H., Suresh, N. L., Rymer, W. Z., and Hu, X. (2018). Relative contribution of endorsed by the publisher. different altered motor unit control to muscle weakness in stroke: a simulation study. J. Neural Eng. 15, 016014. doi: 10.1088/1741-2552/aa925d Copyright © 2021 Koelewijn, Audu, del-Ama, Colucci, Font-Llagunes, Shulga, A., Lioumis, P., Kirveskari, E., Savolainen, S., and Mäkelä, J. P. (2021). A Gogeascoechea, Hnat, Makowski, Moreno, Nandor, Quinn, Reichenbach, Reyes, novel paired associative stimulation protocol with a high-frequency peripheral Sartori, Soekadar, Triolo, Vermehren, Wenger, Yavuz, Fey and Beckerle. This is an component: A review on results in spinal cord injury rehabilitation. Eur. J. open-access article distributed under the terms of the Creative Commons Attribution Neurosci. 53, 3242–3257. doi: 10.1111/ejn.15191 License (CC BY). The use, distribution or reproduction in other forums is permitted, Soekadar, S., Witkowski, M., Gómez, C., Opisso, E., Medina, J., Cortese, M., provided the original author(s) and the copyright owner(s) are credited and that the et al. (2016). Hybrid eeg/eog-based brain/neural hand exoskeleton restores original publication in this journal is cited, in accordance with accepted academic fully independent daily living activities after quadriplegia. Sci. Robot 1, 32–96. practice. No use, distribution or reproduction is permitted which does not comply doi: 10.1126/scirobotics.aag3296 with these terms. Frontiers in Neurorobotics | www.frontiersin.org 8 December 2021 | Volume 15 | Article 750519