Current Reports (2020) 1:131–144 https://doi.org/10.1007/s43154-020-00015-4

REHABILITATION AND ASSISTIVE ROBOTICS (M RAISON AND S ACHICHE, SECTION EDITORS)

Human-Robot Interaction in Rehabilitation and Assistance: a Review

Abolfazl Mohebbi1

Published online: 11 August 2020 # Springer Nature Switzerland AG 2020

Abstract Purpose of Review Research in assistive and rehabilitation robotics is a growing, promising, and challenging field emerged due to various social and medical needs such as aging populations, neuromuscular, and musculoskeletal disorders. Such robots can be used in various day-to-day scenarios or to support motor functionality, training, and rehabilitation. This paper reflects on the human-robot interaction perspective in rehabilitation and assistive robotics and reports on current issues and developments in the field. Recent Findings The survey on the literature reveals that new efforts are put on utilizing machine learning approaches alongside novel developments in sensing technology to adapt the systems with user routines in terms of activities for assistive systems and exercises for rehabilitation devices to fit each user’s need and maximize their effectiveness. Summary A review of recent research and development efforts on human-robot interaction in assistive and rehabilitation robotics is presented in this paper. First, different subdomains in assistive and rehabilitation robotic research are identified, and accord- ingly, a survey on the background and trends of such developments is provided.

Keywords Human-robot interaction . Artificial intelligence . Human-centered design . Feedback control system . Assistive . Rehabilitation robotics

Introduction active prosthetics or exoskeletons for assisting different sen- sorimotor functions such as arm, hand, leg, ankle [3, 4•]. Rehabilitation robotics is a research field focused on augment- With fast-growing technological developments in artificial ing and analyzing rehabilitation procedures by using robotic intelligence, sensing, computation and processing, prototyping, systems. Such systems are developed to aid various methods and fabrication, an increasing number of research efforts have of therapeutic training and assessment of sensorimotor perfor- been recently focused on assistive and rehabilitation robotics due mance [1]. Robotic rehabilitation has been very well received to their vast capabilities in rehabilitating and empowering users. by both patients and clinical professionals and has been found This corresponds to the significant social needs such as motion to be an effective method for motor function therapy in motor aids for the elderly or assisting patients with motor impairments, impairment patients such as [2]. Assistive robotics, on as well as worker augmentation in the industry dealing with the other hand, aims at providing support to patients with intensive physical duties [5, 6]. Moreover, such robotic platforms disabilities to perform their activities of daily living (ADLs) can provide support to users by correcting or preventing unde- with more independence. Examples are moving, grasping, and sired motions, extending the range of motion and workspace, handling objects, eating, etc. Often, a robotic system can be adding extra load-bearing, and incorporating specific force and purposed for both rehabilitation and assisting. For example, motion control capabilities [7]. A majority of previous designs for assistive and rehabilitation robots are functional in terms of delivering the required outputs and performing desired tasks, but This article belongs to the Topical Collection on Rehabilitation and their efficiency is in question due to not using the state-of-the-art Assistive Robotics technologies, and also incorporating insufficient knowledge about the human user during the design and control conception * Abolfazl Mohebbi [8]. Hence, a human-centric approach in design is needed that [email protected] considers various aspects such as the human neural and muscu- 1 Department of Mechanical Engineering, Polytechnique Montréal, loskeletal system and primarily it calls for a focus on human- C.P. 6079, succ. Centre-ville, Montreal, QC H3C 3A7, Canada robot interaction (HRI) and interface technologies [9]. 132 Curr Robot Rep (2020) 1:131–144

Human-robot interaction is currently a very extensive and react and adapt [21]. The assistive robotics, on the other hand, diverse field of research dedicated to understanding, design- cover a broad range of systems, from assistive robots for ma- ing, and evaluating robotic systems for use by or with humans nipulation and mobility that provide support to patients with either in their physical proximity or remotely [10]. An HRI motor function deficits [22], to wearable robots that physically problem is essentially understanding and shaping the interac- augment the body [23], or social assistive robots that aid in tions between humans and robots by assessing the capabilities education and cognitive rehabilitation [24]. of both sides and designing the technologies that form suitable This review paper describes surveys new developments in interactions. This constitutes a multidisciplinary field of re- different categories of rehabilitation and assistive robotics search where various fields such as cognitive sciences, medi- with a focus on state-of-the-art human-robot interaction ac- cine, engineering, and design come together [11]. cording to each category. Various aspects and domains are Robots use the various sensor information to perceive often involved in the development of HRI systems such as humans in the environment they co-exist in. Significant re- design, prototyping, fabrication, sensing technology, commu- search and development efforts exist in the literature on sens- nications, control, etc. Although covering all of these subjects ing components and software for extracting the human kine- is out of the scope of this paper and we intend to mostly cover matics. Depth cameras, stereo-vision devices, infrared, and the soft aspects of HRI in rehabilitation and assistive robotics laser range-finders are examples of the sensing technologies which are the sensory modalities, feedback control ap- used for motion tracking and kinematic assessment of the proaches, and intelligent algorithms. Five search directories human and the environment [12]. The proprioception sensors were utilized in this search strategy and were chosen for their enable the robots to have information over their own move- content and relevance to the HRI approaches in assistive mo- ment and positional state relative to a reference frame. These bile robots, assistive manipulators, robotic prostheses and or- sensing units include encoders, potentiometers, inertial mea- thoses, and rehabilitation robotics. They were from Ei surement units (IMUs), and accelerometers [13]. Moreover, Compendex, Web of Science, PubMed, Proquest, and force and torque sensors are used to measure interaction Science-Direct. Abstract and full texts were then assessed, forces, torques, pressures, and mechanical stress [14]. Tactile and peer reviewed by the author for direct relevance to the sensing units and pressure arrays are used to create a sense of topic and scope of this study. A total of 91 relevant papers touch for the robot with respect to the environment and users were discovered from papers restricted to the last 5 years. [15]. On the communication level, speech recognition algo- Figure 1 shows the distribution of the reviewed papers used rithms are used to interpret human intentions or commands in each category. through speech and vocal signals. In this regard, natural lan- This paper is organized as follows; “Assistive Mobile guage processing (NLP) is concerned with understanding the Robots” describes the HRI in assistive mobile robots such as interactions between computers and human languages smart wheelchairs and walkers; “Assistive Robotic through tools such as neural network architectures and learn- Manipulators” reviews HRI developments for assistive robot- ing algorithms [16]. ic manipulators; “Robotic Prostheses” is dedicated to robotic Based on the therapeutic and clinical techniques, rehabili- prostheses and their interaction with users; “Robotic tation robots are designed to improve and determine the adapt- Exoskeletons” describes current issues and developments of ability level of patients. In the first category of rehabilitation HRI for robotic orthoses and exoskeletons, and finally, robots, simple positional control of extremities is practiced for “Rehabilitation Robots” provides insights about various as- a passive exercise where the robot is used in a positional pects of HRI in rehabilitation robotics. Concluding remarks trajectory mode which does not involve the active participa- are also provided at the end. tion of a patient on either neuromuscular or sensory levels [17]. The second category of such robots incorporates active assisted exercises, in which the robot moves the patient’sex- Assistive Mobile Robots tremity, e.g., leg, arm, etc., along a predetermined trajectory without any force applied to keep the motion on track [18]. In Applying mobile robotic concepts on assistive devices has the third category, active constrained exercises are utilized been an active research area for over two decades and corre- where the robot applies an opposing force if the extremity sponds to two main applications: smart wheelchair systems moves outside of a predefined path or a 3D virtual space and assistive robotic walkers. The wheelchair is the most fre- [19]. The fourth category of rehabilitation robots implements quently used assistive device and its estimated user population the technique of active resistive exercises where the robot is approximately 65 million worldwide [25]. Wheelchairs are opposes the intended movement by the patient [20••]. either manually operated or power-driven. Due to the lack of Finally, the fifth category robots are designed to be used for motor skills, lack of strength, or visual deficiencies, it is hard adaptive exercises, in which the robot provides a previously for a large population of users to operate their wheelchairs unknown dynamic environment to which the subjects must independently. A smart wheelchair is essentially a power Curr Robot Rep (2020) 1:131–144 133

Fig. 1 The distribution of the reviewed papers used in each considered category of HRI in rehabilitation and assistive robotics

wheelchair with a collection of sensors and computer control impairments, several research efforts focused on developing systems. Research on smart wheelchairs has resulted in the frameworks in which the feedback of the environment is pro- development of navigational algorithms for collision avoid- vided to the users or the controller in terms of forces through ance, autonomously transporting the user between locations, haptic interfaces [28, 29]. Some other recent approaches in- and assisted steering of the device through novel human- corporated voice commands and vocal feedbacks to share machine interfaces. The main motivation for this type of re- control decisions with the user such as obstacle avoidance, search is the fact that the commonly used joysticks are not safe approach to objects, navigation through a specific path, always useful especially for the users with a poor level of and learn from these decisions [30]. For patients who cannot neuromuscular capability. Various sensors are integrated into manipulate a standard joystick, many developments have been smart wheelchairs such as ultrasonic, infrared, laser, and cam- carried out to translate users’ face and body gestures and eye eras. Many recent efforts have contributed to the incorporation movements to control commands for the wheelchair through of state-of-the-art technologies into the user interface to in- visual feedbacks [31, 32]. The same approach of gesture clas- crease user autonomy or fully-automate the device’s naviga- sification and recognition is carried out through gathering sur- tion. Figure 2 shows the critical components of a smart face electromyography (EMG) [33, 34] and electroencepha- wheelchair. lography (EEG) [35•] bio-signals. Multimodal sensory inte- In self-navigating smart wheelchairs, the intended destina- gration is also explored in a few research efforts where differ- tions or directions are communicated to the robot through ent information sources such as bio-signals and feedback de- touch-monitors, microphones for speech recognition systems, vices are used in parallel to perform assisted navigation [36]. or assisted joysticks [26, 27]. For users with visual Similar to the HRI approaches for controlling smart wheel- chairs, sensor feedback devices and bio-signal collection frameworks are used alongside learning algorithms for assisted navigation and driving of walkers [37–39]. Besides motor-impaired patients, visually impaired users can also use these smart walkers for safe indoor and outdoor travels.

Assistive Robotic Manipulators

There are essentially millions of patients worldwide with varying degrees of upper limb or lower limb disabilities lim- iting their reach and object handling for ADLs. These patients include individuals with age-related geriatric problems, stroke survivors, muscular dystrophy (MD), spinal cord injury (SCI), multiple sclerosis (MS), amyotrophic lateral sclerosis (ALS), (CP), spinal muscular atrophy (SMA), and oth- er severe motor paralysis [25]. This highlights the need for an Fig. 2 System components and information flow for a smart wheelchair affordable and effective assistive solution to reach objects and system interact with the environment in a safe manner and increase 134 Curr Robot Rep (2020) 1:131–144 patients’ independence. The development of assistive robotic underside of the jaw [47]. Cio et al. built a proof of concept manipulators (ARMs) has been an active research field for for the control of an assistive robotic arm using a low-cost decades that includes wheelchair-mounted and fixed desktop combination of stereovision and eye-tracking using which configurations. the patients were able to successfully reach and grasp an ob- Despite the mainstream industrial robots, ARMs are usual- ject for 92% of the trials [48]. Kæseler et al. used a brain- ly relocated in space and deal with unstructured and dynami- computer interface (BCI) based on steady-state visual evoked cally changing environments to perform various tasks and potentials (SSVEP) for a full workspace control of an ARM interactions such as eating, object grasping and manipulation, [49]. Zeng et al. developed a hybrid HRI using a BCI and a opening doors, or even shaking hands with other people. gaze-tracking interface for continuous modulation of the Hence, in several cases, the user contributes, fully or partially, movement speed via the motor intentions. They also proposed to the control of the ARM, through an interface. Many re- a shared control paradigm that combined user input with an search efforts have focused on developing different user inter- autonomous control algorithm [50]. faces for assistive robotic manipulators to improve the perfor- Besides commanding the ARM to perform specific tasks, it mance of accomplishing functional activities. Furthermore, is equally important to configure a safe human-robot interac- with the development of more powerful embedded computers, tion. Industrial robotic manipulators are usually designed for artificial intelligence is utilized more often to perform object applications that typically require high speed, high accuracy, recognition and path planning to help patients with and heavy payload capabilities. Hence, they are usually con- performing tasks more independently and efficiently. Similar fined within a protected space because it is not considered safe to smart wheelchairs, a sensitive joystick is the most common for human operators to be in their close vicinity. ARMs, on the user interface device for an assistive robotic manipulator. other hand, are near patients to assist them with their activities However, for decades, researchers in academia and industry so they need to be designed to interact with humans safely. have tried to develop more innovative and more inclusive Consequently, they are often restricted in terms of joint and types of interfaces. Voice control and speech recognition sys- end-effector speed, acceleration, and force. Special sensors tems [40], electrooculography-based interfaces [41], eye-gaze and control mechanisms are used to predict, detect, and avoid detection devices [42], vision-based gesture recognition algo- collisions to guarantee the safety of HRI. Figure 3 describes rithms [43], and bio signal-based interfaces [44] are developed various methods for safe human-robot interactions. to translate user intentions to robot commands to perform Sharkawy et al. designed a multilayer feedforward neural specific tasks. In addition to the standard interfaces, various network-based and trained it using data acquired from the approaches were also proposed for intelligent detection and coupled dynamics of the manipulator with and without exter- grasping of objects within the workspace of the ARM [45]. nal contacts to detect unwanted collisions and to identify the Furthermore, a combination of sensing modalities, learning collided link using only the intrinsic joint position and torque methods, and control schemes are used by many researchers to sensors [51]. Li et al. proposed an adaptive compliant control create more efficient HRI systems. Chu et al. created a frame- in which the end-effector’s motions are constrained by human work that used a tongue-driven interface to obtain the user’s arm joint limits. They used an integral reinforcement learning intent and verify it visually through an augmented reality (IRL) to solve a linear quadratic regulation to obtain optimal headset [46]. Sasaki et al. developed a novel tongue interface impedance of the human arm model [52]. Sangiovanni et al. based on the classification of tongue motions from the surface developed a real-time collision avoidance approach for safe EMG signals of the suprahyoid muscles detected at the HRI using deep reinforcement learning applied to a robot with

Fig. 3 Taxonomy of the research methods for safe human-robot interaction of ARMs Curr Robot Rep (2020) 1:131–144 135 an unstructured workspace populated by different obstacles and used it to obtain EMG signals from five different arm [53]. Liu et al. proposed a method based on the Rapidly- muscles. Then, they used a vision-based machine learning Exploring Random Tree Star (RRT*) approach in combina- classification algorithm to categorize muscle activities tion with artificial potential fields (APF) to achieve a high- concerning seven major tasks [56]. Zizoua et al. developed a accuracy motion planning of an assistive arm [54]. strain-gauge matrix sensor to obtain information about upper limb muscle activity from amputees. Moreover, they proposed a method to identify the intentions of the upper limb move- Robotic Prostheses ments [57]. Studies show that lack of natural-feeling sensory feedback A robotic or an active prosthesis is a type rehabil- and improper embodiment are amongst the main reasons the itation robot that is used to replace a lost limb by being at- available robotic prostheses are not well received by patients tached to a patient’s body and replicating its functions [58, 59]. To close the loop of the prosthesis control, a reverse throughout daily activities. Robotic prostheses directly inter- flow of information from the device to the user is crucially act with the human body as their function is usually needed. Figure 4 depicts the active prosthesis control loop commanded and controlled in real-time by the patients via including the interactions between the device and human user. muscle or brain signals. The important traits in designing ro- Besides observing the device in operation, the simplest forms botic prostheses are anthropomorphic appearance, physical of feedback are the vibrations and impact forces the user specifications, interpreting user intent, and mimicking mo- senses on the socket of the limb. In fact, recent studies show tions, force exertion, or grip patterns of the actual human that the sense of muscle vibration from the missing limb body. Recently, many research efforts have been dedicated movement provided patients with better motor control even to developing prostheses that more closely approximate the without visually observing the prosthesis. Marasco et al. pro- capabilities of a lost biological limb. The key to successful posed a method for providing patients with a kinesthetic per- development is to achieve a precise method of realizing the ception of a robotic prosthesis through vibrating the same user intent towards performing a task, and ultimately sensing muscles used for controlling thedeviceusinganeural- the environment while carrying out that task. A combination machine interface [60]. Peripheral intra-neural stimulation of bio-signals can be recorded to form a recognition algorithm through using implantable nerve interfaces can also be a reli- and realize the user intent by the motion and dynamic control able way to provide sensory feedback to upper limb amputees system of the robot. These signals are biometric recordings using a robotic prosthesis. However, it is still necessary to from the electrical activity of the residual limb muscles, brain focus more research efforts on providing encoding strategies activity, or contact forces in a socket. Guo et al. designed a for these stimulations to produce sensations that are natural hybrid human-machine interface for prosthetic control using and effective for prosthesis control. EMG and near-infrared spectroscopy (NIRS) to overcome the Valle et al. proposed a hybrid encoding strategy based on limitations of using EMG only due to noise and sensor cross- the simultaneous biomimetic frequency and amplitude talk [55••]. Pancholi et al. proposed an efficient multichannel neuromodulation to achieve near-natural tactile feedback EMG signal acquisition system for an upper limb prosthesis [61]. They demonstrated that their strategy improved the

Fig. 4 Control interactions between a user and the robotic prosthesis 136 Curr Robot Rep (2020) 1:131–144 control dexterity and accuracy of the prosthesis. Vu et al. used transhumeral amputees based on surface EMG and kinematic a regenerative peripheral nerve interface (RPNI) to serve as a data [69]. biologically stable bio-amplifier to produce electromyography signals with large signal-to-noise ratios. Using these RPNI signals, subjects were able to successfully control a hand pros- Robotic Exoskeletons thesis in real-time for up to 300 days without system recali- bration [62]. Mastinu et al. developed and used an implanted A robotic exoskeleton is an active orthosis device to help pa- neuromusculoskeletal interface built upon osseointegration to tients with mobility and manipulation related disabilities and control their artificial limb. They also designed a controller should be deployable in daily life activities [70]. More specif- that allowed for bioelectric signal acquisition, processing, ically, an upper limb exoskeleton is a wearable robot aimed to decoding of motor intent, prosthetic control, and sensory feed- improve mobility and manipulation of the upper extremities, back. This system included a neurostimulator to provide direct while a lower limb exoskeleton is a powered multi-joint ortho- neural feedback based on sensory information [63]. sis that is able to move leg joints without external support. The The control system of the robotic prosthesis receives com- main goal in designing an upper limb exoskeleton is to deliver plex data patterns from the user and makes real-time motor precision and dexterity in manipulation while the design of a function decisions with regards to the learned prediction of the lower limb exoskeleton is more focused on seamless perfor- user intent. User intended activities are identified through pat- mance in supporting patients in standing and walking. tern recognition methods applied to myoelectric or other bio- Depending on the level of the user motor functionality, the signals. Usually, a classifier is trained to differentiate between control of robotic exoskeletons can range from assistive com- joint actuators of the robotic prosthesis, using patterns from pliant to fully automated. Robotic exoskeletons can be also multi-channel EMG data. Research efforts on this subject are used as an augmentation device for healthy users to increase mainly focused on improving the interpretation of myoelectric their strength, speed, or endurance for various industrial or patterns, and the simultaneous pattern recognition and control military applications. However, this category of exoskeletons of multiple functions to increase the robustness of the tasks is out of the scope of the present review. carried out by the device. Essentially, tasks from activities of Similar to robotic prostheses, an essential component in the daily living include simultaneous movements of multiple de- control of robotic exoskeletons is the acquisition and identifi- grees of freedom; hence combined joint motions must be sep- cation of human intent which is carried out by different means arately classified. Samuel et al. proposed a pattern recognition of human-robot interaction and acts as an input to the control method based on three new time-domain features to improve system. A cognitive human-robot interaction (cHRI) uses the performance of EMG classification in arm movement for EEG signals from the central nervous system (CNS) to the an active prosthesis [64]. Additionally, they investigated the musculoskeletal system, or surface EMG signals, to identify effect of mobility on the performance of the motion classifier the human intent before the occurrence of any actual body based on EMG and accelerometer signals. Ultimately, they motion of the user and then estimate the required torque or proposed a method to mitigate such effects on the motion position inputs [71••]. Figure 5 a illustrates the formation of a classifier using three different classifiers including linear dis- torque estimation and torque control loop using a cognitive criminant analysis (LDA), k-nearest neighbors (kNN), and human-robot interaction mechanism. In some cases, it is also support vector machine (SVM) [65]. possible to use brain activity to provide high-level commands During the past few years, deep learning approaches are to the device, which is then shared with the control system to utilized as a new tool to perform classification and regression ensure safe execution. Shared control strategies reduce the tasks directly from high-dimensional raw EMG data, without cognitive workload, as the users are not concerned with mid- determining and identifying any data features [66]. Atzori to-low-level execution over long periods or during complex et al. designed a deep convolutional neural network (CNN) operations [72]. Figure 5 b shows a shared control paradigm for the classification of an average of 50 hand movements for a robotic exoskeleton. through EMG data and use the outcome to control a robotic Gordleeva et al. proposed a human-robot interface for as hand [67]. Dantas et al. presented four decoding methods for lower limb exoskeleton which records and processes multi- the recognition of movement intent from surface EMG signals modal signals collected using a foot motor imagery-based using polynomial Kalman filters (KFs), multilayer perceptron brain-machine interface (BMI) and multichannel EMG signals (MLP) networks, convolutional neural networks (CNN), and recorded from leg muscles [73]. They essentially integrated long short-term memory (LSTM) networks [68]. They vali- EEG and EMG modalities, to provide real-time control of the dated their work through its application on a problem of pros- exoskeleton. They proposed an estimation method for the thesis control. Gaudet et al. used a multi-layer perceptron neu- intended joint angles of a lwer-limb exoskeleton from EMG ral network with backpropagation training algorithm to accu- signals using a multi-layer backpropagation neural network rately classify upper limb phantom movements in (BPN). They specified mean absolute value (MAV) and root Curr Robot Rep (2020) 1:131–144 137

Fig. 5 Lower limb exoskeleton control scheme. a Joint torque control system based on cognitive human-robot interactions. b Real- time shared control of a lower limb exoskeleton using brain- machine interface (BMI) and electromyography signals

mean square (RMS) of the signals as features used in their signals using the spectral power correlation (SPC) to classify classification algorithm [74]. Gui et al. presented an adaptive the grasp attempt and resting states of the user [79]. Accogli method to estimate active joint torque using EMG signals for a et al. developed a fast and advanced machine learning (ML) lower limb exoskeleton with two degrees of freedom. Their algorithm to control an upper limb exoskeleton based on sup- estimation method used a radial basis function neural network port vector machines (SVMs) for detecting the motion inten- (RBFNN) to form an extended nonlinear controller which tion and decoding the intended movement direction [80]. In an eliminated the need for the calibration of the EMG to torque effort by Irastorza-Landa et al., healthy subjects were estimator model [75]. Beil et al. proposed a movement classi- employed to perform reaching movements in four directions fication system for a lower limb exoskeleton based on a hid- and five different hand movements by wearing an exoskeleton den Markov model (HMM), which enables the online classi- device while EMG signals were recorded form their muscles. fication of multi-modal sensor data acquired from 3D-force Moreover, they developed an offline classifier based on a sensors and IMUs [76]. back-propagation artificial neural network (ANN) trained Research efforts on upper limb exoskeleton are focused on with the waveform length as time-domain feature extracted delivering interfaces and decoding approaches to provide pre- from EMG signals to classify discrete movements [81]. cision and dexterity in a wider range of motions compared to A physical human-robot interaction (pHRI) uses measure- lower limb exoskeletons. Machine-learning algorithms are ments of the force or changes in the joint positions due to the particularly useful for detecting the user’s motion intentions motion of the musculoskeletal system, as control inputs to the based on classified bio-signals and applied for real-time con- exoskeleton robot. In such cases, the control system of the robot trol of such devices [77]. Trigili et al. proposed a method using aims at minimizing the human efforts needed to accomplish the EMG signals from seven upper limb muscles and incorporated tasks, hence, to get a compliant behavior. More precisely, min- a Gaussian mixture model (GMM) to identify the onset of the imum interaction forces are desirable while task tracking errors movement in different scenarios. They used time-domain fea- should be eliminated. To get a compliant controller, the inter- tures extracted from single muscles and also multiple muscles action forces are usually controlled using either the impedance [78]. Chowdhury et al. created a hybrid interface device for or admittance controllers [82, 83]. In order to model the human- controlling a hand exoskeleton by combining EEG and EMG robot interaction, these controllers use a virtual impedance term 138 Curr Robot Rep (2020) 1:131–144 for the manipulator and human limb impedance. Figure 6 a torque and returns the position. This control scheme com- illustrates a single-degree-of-freedom linear model of a coupled prises the admittance model and a position control loop. The human-exoskeleton system. admittance model receives torques and produces joint posi- In an impedance control scheme, the position or velocity is tions and then a position controller regulates the joint angles of the system input while force or torque is the output. In this the exoskeleton based on the position references yielded from type of interaction control, the force produced by the patient the admittance model. The appropriate values for the virtual corresponds to how fast the robot should move. Basically, this mass, damper, and spring coefficients should be tuned and controller not only controls the position and the force, but also specified to make the admittance control adapt to the interac- the relationship between them, hence the interaction between tion with the patient’s limb. Figure 6 c describes the joint the exoskeleton and the human body. Figure 6 bshowsan admittance control system for a lower limb exoskeleton. impedance control system for a lower limb exoskeleton. The Both impedance and admittance control have been widely impedance model receives the position error with respect to used in the applications of pHRI for both upper limb and each joint and generates the torque values which then are fed lower limb robotic exoskeletons. Usually, the impedance to a torque controller that minimizes the error between the model parameters are fixed in design after specifications. torques exerted by the exoskeleton and the torque references. However, in some cases, these parameters need to adapt to In the case of using an impedance control system, the robotic the variability of different physical conditions of patients. To dynamics and the interaction dynamics are most probably re- tackle this problem, Alquadi et al. and Li et al. proposed adap- quired. The admittance control is the opposite of the imped- tive methods by incorporating biological signals and using ance control as it reacts with the robot’smotions;itgetsthe neural networks to compensate for the unknown nonlinear

Fig. 6 Joint impedance control model for a lower limb robotic exoskeleton. a 1-DOF linear model of a coupled human- exoskeleton system. The term τp represents the net torque exerted by the exoskeleton on the human limb at the interaction point p, τh is the net muscle torque assisted by the exoskeleton, and θ is the joint position. b Joint impedance control model. c Joint admittance control model Curr Robot Rep (2020) 1:131–144 139 robot dynamics [84, 85]. Figueiredo et al. used an online stiff- information provided to the patient about the procedure and ness controller to ensure the adaptive behavior of the robot progress of the therapy, and the closed-loop control system through analyzing the interaction torque-angle characteristics dealing with motion planning and compliant behavior of the from experimental data [86]. In a research work by Li et al., a robot corresponding to patient’smotion.Figure7 bshowsan new adaptive impedance control was proposed in which an example of a lower limb rehabilitation robot and different integral reinforcement learning (IRL) method was used to interaction modalities involved. solve a linear quadratic regulation and minimize the position Information about the patient’s movement and perfor- tracking errors. They claimed they had obtained optimal im- mance can be collected using different sensing devices such pedance parameters of the human arm with little information as muscle activity signals, musculoskeletal motion trackers, about the musculoskeletal model [52]. Geoffroy et al. pro- position, and force sensors. Serpelloni et al. developed a ro- posed a nonlinear model-predictive control strategy for torque botic system for hand rehabilitation, driven by surface EMG regulation of a complex variable stiffness robot based on var- measurements, based on the mirroring of healthy hand move- iable compliance [87]. Azimi proposed an adaptive imped- ments [91]. Sarasola-Sanz et al. presented a control method ance controller to increase the robot’srobustnesstovariations for a 7-DOF upper limb rehabilitation robot using EEG, EMG, of ground reaction forces. They used a particle swarm opti- and hybrid brain-machine interfaces. Their data acquisition mizer (PSO) to find the optimal design parameters of the con- system incorporated central and peripheral structures of the troller and the adaptation law [88]. Torabi et al. proposed an nervous system in a biologically inspired hierarchical control admittance controller based on a nonlinear adaptive sliding scheme [92]. Liu et al. developed an upper limb rehabilitation mode scheme for a lower limb exoskeleton robot to provide robot for at-home use by children with cerebral palsy [93]. robustness against disturbances while ensuring a compliant Their device includes portable wearable sensing using com- behavior during patient-exoskeleton interaction [89]. prised accelerometers and surface electromyography sensors to capture functional movements. The feedback towards the exercise and the progress of the Rehabilitation Robots interventions are usually delivered to the patients by monitors, virtual reality headsets, haptic feedback devices, and neuro- Rehabilitation robots are now able to complement existing stimulation. Calabro et al. studied the effect of using a virtual clinical practices to help patients recovering from a wide range reality (VR) system integrated with a robot-assisted gait train- of neurological and coordination disorders such as stroke, ing for MS patients with walking disabilities [94]. They brain injury, spinal cord injury, cerebral palsy (children), or- showed that using a VR environment is particularly useful thopedic surgeries, etc. Studies have shown that using reha- as it trains different brain areas involved in motor planning bilitation robots for physical therapy are long-lasting and more and learning. Archambault et al. proposed a rehabilitation in- efficient compared to conventional therapy methods [90]. tervention for upper extremity by integrating an upper limb Based on the principles of , recovery for such rehabilitation robot, a virtual reality system, and neuromuscu- patients is highly related to the intensity and frequency of lar electrical stimulation (NMES) [95]. Berezny et al. devel- therapeutic intervention—the concept of repetitive patterning. oped a lower limb rehabilitation robot for acute, bed-bound However, not all patients are able to exercise as frequently as patients by incorporating virtual reality and haptic feedback to need due to the lack of physiotherapy resources. increase user engagement [96]. Ocampo et al. integrated an Rehabilitation robots can tackle this problem as they optimize augmented reality (AR) game environment into an upper limb neurological recovery by providing increased, flexible, and robotic rehabilitation system to provide colocation between customizable access to therapy exercises. They reduce the cost visual and haptic feedback [97]. of the treatment and also the workload of the physical thera- The control system of the rehabilitation robots should be pists, provide more extensive programs and suggest new in- customizable in terms of altering the amount of assistance it sights into the treatment process. Moreover, these devices can provides to the patient to reduce the reliance on the therapist. collect quantitative information about the performance of the Like robotic exoskeletons, the impedance-based control sys- patient in the therapeutic exercise and track their progress. tems are usually used to regulate the interaction between the The human-robot interaction aspect of the rehabilitation robot and patient in order to tailor the intervention based on the robotics usually involves three types of modalities: therapist- user’s condition and also progress. Gui et al. developed a gait robot interaction, patient-robot interaction, and robot learning rehabilitation robot with multiple gait patterns, which can be (Fig. 7a). The therapist-robot interactions often comprised the controlled by the active motion intention of users [98]. They monitoring, customization, and parameter control of the ther- created a multimodal cognitive-physical interaction system to apeutic intervention. The patient-robot interaction aspect in- better engage subjects during gait rehabilitation by adopting a cludes the information collected from the patient in terms of brain-computer interface and EMG-based admittance control. performance assessment and control system feedback, the Jamwal et al. proposed an interactive intervention based on the 140 Curr Robot Rep (2020) 1:131–144

Fig. 7 a Category human-robot interaction modalities for rehabilitation robotics. b An example of human-robot interaction modalities in a lower limb rehabilitation robotic system

impedance control for a lightweight ankle rehabilitation robot. restorations of lost or diminished functions [101]. These This training program allowed modification of the robot- algorithms exploit co-adaptation between the robot and imposed motions according to the patient’s level of disability patients. It is possible to use training data formed by syn- [99]. Song et al. presented an impedance model for the dynam- thetic patterns of kinematic features and bio-signal activa- ics of the post-stroke patient arm and proposed a control tion, e.g., EMG, based on a priori-known motor function scheme which included an adaptive damping control strategy, and use the robotic therapist to adapt the patient to them. to robustly control an upper limb rehabilitation robot moving Machine learning methods have been used for the classifi- along specific trajectories. They used an adaptive window cation of patients’ motor functions and their health condi- method with least mean square for the on-line estimation of tions. Furthermore, these algorithms can be incorporated to the impedance parameters of the arm [100]. build intelligent rehabilitation systems suggesting inter- Besides monitoring the therapeutic intervention and quan- ventions based on patient conditions and adjusting assess- tifying the patient’s performance and progress, designing in- ment methods [102•]. telligent algorithms for detailed analysis of the subjects Dolatabadi et al. used machine learning methods alongside and their adaptability can lead to faster and more complete an affordable sensing approach to differentiate between Curr Robot Rep (2020) 1:131–144 141 healthy and pathological gait patterns resulting from stroke or field of HRI for assistive and rehabilitation robotics. brain injuries [103]. They acquired gait features such as the The interaction modalities in assistive mobile robotics, orientations of the hip, spine, shoulders, neck, knee, and ankle assistive robotic manipulators, robotic prosthesis, robotic during walking using a motion tracking system, and used a exoskeletons, and rehabilitation robotics have been Gaussian process latent variable model (GPLVM) to analyze discussed and our perspective on the HRI components the features and identify the role of each body part in separating of such systems and also challenges facing them have pathological patterns from healthy patterns. Cui et al. devel- been presented from the standpoints of sensory modali- oped an intelligent gait analysis system for the identification ties and data fusion, feedback control approaches, and and assessment of gait abnormality amongst post-stroke intelligent algorithms. Some future research directions hemiparetic patients [104]. They recorded marker trajectory, for HRI in rehabilitation and assistance can include: ground reaction forces, and EMG signals, from subjects during walking and used machine learning methods on the multimod- & The physical implications of long-term exposure and use al sensor fusion data to distinguish the hemiparetic gait patterns of rehabilitation and assistive robots from healthy patterns. Badesa et al. presented an adaptive & The HRI in parallel and cooperative robots used for assis- robot-aided therapy using machine learning classification to tance and rehabilitation dynamically apply modifications to the intervention and the & Development of more advance input devices benefiting provided feedback to the subject in compliance with the spe- from multimodal bio-signals cific state of each patient [105]. Barzilay et al. designed an & Dexterous manipulation and motion control of redundant analytical neural network (ANN) and trained it using upper assistive and rehabilitation robots limb kinematics and electromyography (EMG) signals to ad- & Development of artificial intelligence-based adaptive im- just the difficulty level of an upper limb robotic rehabilitation pedance control for physical HRI task [106]. Cai et al. used machine learning techniques for real- & Design shared control and mutual adaptation paradigms time monitoring of possible compensations in post-stroke pa- by using machine learning techniques tients by using pressure distribution data [107]. Compensations & Using intelligent algorithms for learning the actions and are often carried out by stroke patients during rehabilitation interventions particularly demonstrated by the therapists. interventions, leading to a less desirable and longer recovery. The proposed method used a support vector machine (SVM) to Compliance with Ethical Standards classify the compensatory patterns online and accordingly pro- vide an assistive force to patients to reduce compensations Conflict of Interest The author declares that he has no conflict of through a rehabilitation robot. interest.

Human and Animal Rights and Informed Consent This article does not contain any studies with human or animal subjects performed by any of Conclusion and Future Directions the authors.

During past decades, significant research and develop- ment efforts have been focused on assistive and rehabil- References itation robots to aid patients with various sensorimotor, musculoskeletal, and central nervous system deficits. As Papers of particular interest, published recently, have been the need for assistive and rehabilitation systems grow, highlighted as: and due to existing human in the loop, the design, con- • Of importance trol, sensing, and optimizing such systems are becoming •• Of major importance more complex with recent technological advancements in artificial intelligence, computational power, and fab- 1. Marchal-Crespo L, Reinkensmeyer DJ. Review of control strate- rication. Human-robot interaction plays a crucial role in gies for robotic movement training after neurologic injury. J effectively incorporating robots in assistance and reha- Neuroeng Rehab. 2009;6(1):20. bilitation scenarios. Three types of interactions have 2. Krebs HI, Palazzolo JJ, Dipietro L, Ferraro M, Krol J, Rannekleiv K, et al. Rehabilitation robotics: performance-based progressive been identified in the literature for rehabilitation and robot-assisted therapy. Auton Robot. 2003;15(1):7–20. assistive robots: social, cognitive, and physical. From 3. Leonardis D, Barsotti M, Loconsole C, Solazzi M, Troncossi M, the therapeutic standpoint, these robots should possess Mazzotti C, et al. An EMG-controlled robotic hand exoskeleton multiple interaction modes, ranging from passive to for bilateral rehabilitation. IEEE Trans Haptics. 2015;8(2):140– 51. active-assisted and active-resisted movements. In this 4.• Lebrasseur A, Lettre J, Routhier F, Archambault PS, Campeau-Lecours paper, we have presented a comprehensive review of A. Assistive robotic arm: evaluation of the performance of intelligent the current developments and the state-of-the-art in the algorithms. Assist Technol. 2019:1–10 This paper provides 142 Curr Robot Rep (2020) 1:131–144

comparative insights on how an intelligent algorithm should Recent Developments in Control, Automation & Power perfom when embedded in the control system of an assistive Engineering (RDCAPE): IEEE; 2019. robotic manipulator. 24. Abdi J, Al-Hindawi A, Ng T, Vizcaychipi MP. Scoping review on 5. Windrich M, Grimmer M, Christ O, Rinderknecht S, Beckerle P. the use of socially assistive robot technology in elderly care. BMJ Active lower limb prosthetics: a systematic review of design is- Open. 2018;8(2). sues and solutions. Biomed Eng Online. 2016;15(3):140. 25. Organization WH. Guidelines on the provision of manual wheel- 6. Huysamen K, de Looze M, Bosch T, Ortiz J, Toxiri S, O'Sullivan chairs in less resourced settings. 2008. LW. Assessment of an active industrial exoskeleton to aid dynam- 26. Ghorbel M, Pineau J, Gourdeau R, Javdani S, Srinivasa S. A ic lifting and lowering manual handling tasks. Appl Ergon. decision-theoretic approach for the collaborative control of a smart 2018;68:125–31. wheelchair. Int J Soc Robot. 2018;10(1):131–45. 7. Shi L, Yu Y, Xiao N, Gan D. Biologically inspired and rehabili- 27. Schwesinger D, Shariati A, Montella C, Spletzer J. A smart wheel- tation robotics. Appl Bionics Biomech. 2019;2019:1–2. chair ecosystem for autonomous navigation in urban environ- 8. Yan T, Cempini M, Oddo CM, Vitiello N. Review of assistive ments. Auton Robot. 2017;41(3):519–38. strategies in powered lower-limb orthoses and exoskeletons. 28. Devigne L, Pasteau F, Babel M, Narayanan VK, Guegan S, Robot Auton Syst. 2015;64:120–36. Gallien P, editors. Design of a haptic guidance solution for 9. Beckerle P, Christ O, Wojtusch J, Schuy J, Wolff K, Rinderknecht assisted power wheelchair navigation. 2018 IEEE International S, et al., editors. Design and control of a robot for the assessment Conference on Systems, Man, and Cybernetics (SMC): IEEE; of psychological factors in prosthetic development. 2012 IEEE 2018. International Conference on Systems,Man,andCybernetics 29. Chuy OY, Herrero J, Al-Selwadi A, Mooers A. Control and eval- (SMC): IEEE; 2012. uation of a motorized attendant wheelchair with haptic interface. J 10. Adams JA, editor. Human-robot interaction design: understanding Med Dev. 2019;13(1). user needs and requirements. Proceedings of the Human Factors 30. MSI S, Nordin S, Ali AM, editors. Voice control intelligent wheel- and Ergonomics Society Annual Meeting. Los Angeles, CA: chair movement using CNNs. 2019 1st International Conference SAGE Publications Sage CA; 2005. on Artificial Intelligence and Data Sciences (AiDAS): IEEE; 11. Goodrich MA, Schultz AC. Human-robot interaction: a survey. 2019. Found Trends Hum-Comput Interac. 2007;1(3):203–75. 31. Rabhi Y, Mrabet M, Fnaiech F. Intelligent control wheelchair 12. Zheng Y, Zhong P, Liu K, Yang K, Yue Q. Human motion cap- using a new visual joystick. J Healthcare Eng. 2018;2018:1–20. ture system based 3D reconstruction on rehabilitation assistance 32. Rabhi Y, Mrabet M, Fnaiech F. A facial expression controlled stability of lower limb exoskeleton robot climbing upstairs pos- wheelchair for people with disabilities. Comput Methods Prog ture. IEEE Sensors J. 2019. Biomed. 2018;165:89–105. 13. Niku SB. Introduction to robotics: analysis, control, applications: 33. Rakasena E, Herdiman L, editors. Electric wheelchair with John Wiley & Sons; 2020. forward-reverse control using electromyography (EMG) control 14. Lynch KM. Park FC. Modern robotics: Cambridge University of arm muscle. Journal of Physics: Conference Series; 2020. Press; 2017. 34. Kumar B, Paul Y, Jaswal RA, editors. Development of EMG 15. Luo S, Bimbo J, Dahiya R, Liu H. Robotic tactile perception of controlled electric wheelchair using SVM and kNN classifier for object properties: a review. Mechatronics. 2017;48:54–67. SCI patients. International Conference on Advanced Informatics 16. Foster ME. Natural language generation for social robotics: op- for Computing Research: Springer; 2019. portunities and challenges. Philos Trans R Soc B. 35.• Zgallai W, Brown JT, Ibrahim A, Mahmood F, Mohammad K, 2019;374(1771):20180027. Khalfan M, et al., editors. Deep learning AI application to an 17. Pan L, Song A, Duan S, Yu Z. Patient-centered robot-aided pas- EEG driven BCI smart wheelchair. 2019 Advances in Science sive neurorehabilitation exercise based on safety-motion decision- and Engineering Technology International Conferences (ASET): making mechanism. Biomed Res Int. 2017;2017:1–11. IEEE; 2019. This paper is a very good and detailed example of 18. Colombo R, editor. Robot assisted exercise: modelling the recov- using Deep Learning approaches to form an intelligent assis- ery process to personalise therapy. converging clinical and engi- tive control system using bio-signals. neering research on neurorehabilitation III: Proceedings of the 4th 36. Coelho FJdOR. Multimodal interface for an intelligent wheel- International Conference on NeuroRehabilitation (ICNR2018), chair. 2019. October 16–20, 2018, Pisa, Italy: Springer; 2018. 37. Wachaja A, Agarwal P, Zink M, Adame MR, Möller K, Burgard 19. Becker S, Bergamo F, Williams S, Disselhorst-Klug C. W. Navigating blind people with walking impairments using a Comparison of muscular activity and movement performance in smart walker. Auton Robot. 2017;41(3):555–73. robot-assisted and freely performed exercises. IEEE Trans Neural 38. Alves J, Seabra E, Caetano I, Gonçalves J, Serra J, Martins M, Syst Rehab Eng. 2018;27(1):43–50. et al., editors. Considerations and mechanical modifications on a 20.•• Akdogan E, Aktan ME. Impedance control applications in therapeutic smart walker. 2016 International Conference on Autonomous exercise robots. Control Systems Design of Bio-Robotics and Robot Systems and Competitions (ICARSC): IEEE; 2016. Bio-mechatronics with Advanced Applications: Elsevier; 2020. 39. Caetano I, Alves J, Gonçalves J, Martins M, Santos CP, editors. p. 395–443. This paper explains a very common and essential Development of a biofeedback approach using body tracking with HRI contro strategy which is impedance control, applied to active depth sensor in ASBGo smart walker. 2016 International important therapeutic robotic exercises. Conference on Autonomous Robot Systems and Competitions 21. Michmizos KP, Krebs HI. Pediatric robotic rehabilitation: current (ICARSC): IEEE; 2016. knowledge and future trends in treating children with sensorimo- 40. Poirier S, Routhier F, Campeau-Lecours A, editors. Voice control tor impairments. NeuroRehabilitation. 2017;41(1):69–76. interface prototype for assistive robots for people living with upper 22. Campeau-Lecours A, Lamontagne H, Latour S, Fauteux P, Maheu limb disabilities. 2019 IEEE 16th International Conference on V, Boucher F, et al. Kinova modular robot arms for service robot- Rehabilitation Robotics (ICORR): IEEE; 2019. ics applications. Rapid Automation: Concepts, Methodologies, 41. Sun L, Sa W, Chen H, Chen Y, editors. A novel human computer Tools, and Applications. IGI Global. 2019:693–719. interface based on electrooculogram signal for smart assistive ro- 23. Kumar V, Hote YV, Jain S, editors. Review of exoskeleton: his- bots. 2018 IEEE International Conference on Mechatronics and tory, design and control. 2019 3rd International Conference on Automation (ICMA): IEEE; 2018. Curr Robot Rep (2020) 1:131–144 143

42. Leroux M, Raison M, Adadja T, Achiche S, editors. Combination 58. Wijk U, Carlsson I. Forearm amputees' views of prosthesis use of eyetracking and computer vision for robotics control. 2015 and sensory feedback. J Hand Ther. 2015;28(3):269–78. IEEE International Conference on Technologies for Practical 59. Graczyk EL, Schiefer MA, Saal HP, Delhaye BP, Bensmaia SJ, Robot Applications (TePRA): IEEE; 2015. Tyler DJ. The neural basis of perceived intensity in natural and 43. Haseeb MA, Kyrarini M, Jiang S, Ristic-Durrant D, Gräser A, artificial touch. Sci Transl Med. 2016;8(362):362ra142-362ra142. editors. Head gesture-based control for assistive robots. 60. Marasco PD, Hebert JS, Sensinger JW, Shell CE, Schofield JS, Proceedings of the 11th PErvasive Technologies Related to Thumser ZC, et al. Illusory movement perception improves motor Assistive Environments Conference; 2018. control for prosthetic hands. Sci Transl Med. 2018;10(432). 44. Schabron B, Reust A, Desai J, Yihun Y, editors. Integration of 61. Valle G, Mazzoni A, Iberite F, D’Anna E, Strauss I, Granata G, forearm sEMG signals with IMU sensors for trajectory planning et al. Biomimetic intraneural sensory feedback enhances sensation and control of assistive robotic arm. 2019 41st Annual naturalness, tactile sensitivity, and manual dexterity in a bidirec- International Conference of the IEEE Engineering in Medicine tional prosthesis. Neuron. 2018;100(1):37–45 e7. and Biology Society (EMBC): IEEE; 2019. 62. Vu PP, Vaskov AK, Irwin ZT, Henning PT, Lueders DR, Laidlaw 45. Bousquet-Jette C, Achiche S, Beaini D, Cio YL-K, Leblond- AT, et al. A regenerative peripheral nerve interface allows real- Ménard C, Raison M. Fast scene analysis using vision and artifi- time control of an artificial hand in upper limb amputees. Sci cial intelligence for object prehension by an assistive robot. Eng Transl Med. 2020;12(533). – Appl Artif Intell. 2017;63:33 44. 63. Mastinu E, Doguet P, Botquin Y, Håkansson B, Ortiz-Catalan M. 46. Chu F-J, Xu R, Zhang Z, Vela PA, Ghovanloo M, editors. The Embedded system for prosthetic control using implanted neuro- helping hand: an assistive manipulation framework using aug- muscular interfaces accessed via an osseointegrated implant. IEEE mented reality and tongue-drive interfaces. 2018 40th Annual Trans Biomed Circ Syst. 2017;11(4):867–77. International Conference of the IEEE Engineering in Medicine 64. Samuel OW, Zhou H, Li X, Wang H, Zhang H, Sangaiah AK, and Biology Society (EMBC): IEEE; 2018. et al. Pattern recognition of electromyography signals based on 47. Sasaki M, Onishi K, Stefanov D, Kamata K, Nakayama A, novel time domain features for amputees' limb motion classifica- Yoshikawa M, et al. Tongue interface based on surface EMG tion. Comput Electr Eng. 2018;67:646–55. signals of suprahyoid muscles. Robomech J. 2016;3(1):9. 65. Samuel OW, Li X, Geng Y, Asogbon MG, Fang P, Huang Z, et al. 48. Cio Y-SL-K, Raison M, Ménard CL, Achiche S. Proof of concept Resolving the adverse impact of mobility on myoelectric pattern of an assistive robotic arm control using artificial stereovision and recognition in upper-limb multifunctional prostheses. Comput eye-tracking. IEEE Trans Neural Syst Rehab Eng. 2019;27(12): Biol Med. 2017;90:76–87. 2344–52. 66. Ameri A, Akhaee MA, Scheme E, Englehart K. Real-time, simul- 49. Kæseler RL, Leerskov K, Struijk LA, Dremstrup K, Jochumsen taneous myoelectric control using a convolutional neural network. M, editors. Designing a brain computer interface for control of an PLoS One. 2018;13(9). assistive robotic manipulator using steady state visually evoked 67. Atzori M, Cognolato M, Müller H. Deep learning with potentials. 2019 IEEE 16th International Conference on convolutional neural networks applied to electromyography data: Rehabilitation Robotics (ICORR): IEEE; 2019. a resource for the classification of movements for prosthetic 50. ZengH,ShenY,HuX,SongA,XuB,LiH,etal.Semi- hands. Front Neurorobot. 2016;10:9. autonomous robotic arm reaching with hybrid gaze–brain ma- chine interface. Front Neurorobot. 2020;13:111. 68. Dantas H, Warren DJ, Wendelken SM, Davis TS, Clark GA, 51. Sharkawy A-N, Koustoumpardis PN, Aspragathos N. Human– Mathews VJ. Deep learning movement intent decoders trained – with dataset aggregation for prosthetic limb control. IEEE Trans robot collisions detection for safe human robot interaction using – one multi-input–output neural network. Soft Comput. 2019:1–33. Biomed Eng. 2019;66(11):3192 203. 52. Li Z, Liu J, Huang Z, Peng Y, Pu H, Ding L. Adaptive impedance 69. Gaudet G, Raison M, Achiche S. Classification of upper limb control of human–robot cooperation using reinforcement learning. phantom movements in transhumeral amputees using electromyo- IEEE Trans Ind Electron. 2017;64(10):8013–22. graphic and kinematic features. Eng Appl Artif Intell. 2018;68: – 53. Sangiovanni B, Rendiniello A, Incremona GP, Ferrara A, Piastra 153 64. M, editors. Deep reinforcement learning for collision avoidance of 70. Van der Loos HM, Reinkensmeyer DJ, Guglielmelli E. robotic manipulators. 2018 European Control Conference (ECC): Rehabilitation and health care robotics. Springer handbook of – IEEE; 2018. robotics. Springer; 2016. p. 1685 1728. 54. Liu Z, Ai Q, Liu Y, Zuo J, Zhang X, Meng W, et al., editors. An 71.•• Gull MA, Bai S, Bak T. A review on design of upper limb exoskeletons. optimal motion planning method of 7-DOF robotic arm for upper Robotics. 2020;9(1):16 This paper provides important insights limb movement assistance. 2019 IEEE/ASME International about designing upperlimb robotic prosthetics considering their Conference on Advanced Intelligent Mechatronics (AIM): IEEE; interactions with the users, and describes multiple examples with 2019. a critical approach. 55.•• Guo W, Sheng X, Liu H, Zhu X. Toward an enhanced human– 72. Tucker MR, Olivier J, Pagel A, Bleuler H, Bouri M, Lambercy O, machine et al. Control strategies for active lower extremity prosthetics and interface for upper-limb prosthesis control with combined EMG orthotics: a review. J Neuroeng Rehab. 2015;12(1):1. and NIRS signals. IEEE Trans Hum-Mach Syst. 2017;47(4):564–75 73. Gordleeva SY, Lobov SA, Grigorev NA, Savosenkov AO, The research carried out in this paper can be a guide on how to Shamshin MO, Lukoyanov MV, et al. Real-time EEG–EMG hu- implement and enhance bio-signal based human–machine man–machine interface-based control system for a lower-limb interfaces. exoskeleton. IEEE Access. 2020. 56. Pancholi S, Joshi AM. Portable EMG data acquisition module for 74. Dhindsa IS, Agarwal R, Ryait HS, editors. Joint angle prediction upper limb prosthesis application. IEEE Sensors J. 2018;18(8): from Emg signals for lower limb exoskeleton. International 3436–43. Conference and Youth School on Information Technology and 57. Zizoua C, Raison M, Boukhenous S, Attari M, Achiche S. Nanotechnology (ITNT-2016); 2016; Samara. Development of a bracelet with strain-gauge matrix for movement 75. Gui K, Liu H, Zhang D. A practical and adaptive method to intention identification in traumatic amputees. IEEE Sensors J. achieve EMG-based torque estimation for a robotic exoskeleton. 2017;17(8):2464–71. IEEE/ASME Trans Mechatron. 2019;24(2):483–94. 144 Curr Robot Rep (2020) 1:131–144

76. Beil J, Ehrenberger I, Scherer C, Mandery C, Asfour T, editors. rehabilitation of stroke patients. 2017 International Conference Human motion classification based on multi-modal sensor data for on Rehabilitation Robotics (ICORR): IEEE; 2017. lower limb exoskeletons. 2018 IEEE/RSJ International 93. Liu L, Chen X, Lu Z, Cao S, Wu D, Zhang X. Development of an Conference on Intelligent Robots and Systems (IROS): IEEE; EMG-ACC-based upper limb rehabilitation training system. IEEE 2018. Trans Neural Syst Rehab Eng. 2016;25(3):244–53. 77. Barron O, Raison M, Achiche S. Control of transhumeral prosthe- 94. Calabrò RS, Russo M, Naro A, De Luca R, Leo A, Tomasello P, ses based on electromyography pattern recognition: from ampu- et al. Robotic gait training in multiple sclerosis rehabilitation: can tees to deep learning. Powered Prostheses: Elsevier; 2020. p. 1– virtual reality make the difference? Findings from a randomized 21. controlled trial. J Neurol Sci. 2017;377:25–30. 78. Trigili E, Grazi L, Crea S, Accogli A, Carpaneto J, Micera S, et al. 95. Archambault PS, Norouzi-Gheidari N, Kairy D, Levin MF, Milot Detection of movement onset using EMG signals for upper-limb M-H, Monte-Silva K, et al., editors. Upper extremity intervention exoskeletons in reaching tasks. J Neuroeng Rehab. 2019;16(1):45. for stroke combining virtual reality, robotics and electrical stimu- 79. Chowdhury A, Raza H, Dutta A, Prasad G. EEG-EMG based lation. 2019 International Conference on Virtual Rehabilitation hybrid brain computer interface for triggering hand exoskeleton (ICVR): IEEE; 2019. for neuro-rehabilitation. Proceedings of the Advances in Robotics; – 96. Berezny N, Dowlatshahi D, Ahmadi M, editors. Interaction con- 2017. p. 1 6. trol and haptic feedback for a lower-limb rehabilitation robot with 80. Accogli A, Grazi L, Crea S, Panarese A, Carpaneto J, Vitiello N, ’ virtual environments. Proceedings of the 6th International et al. EMG-based detection of user s intentions for human- Conference of Control, Dynamic Systems, and Robotics; 2019. machine shared control of an assistive upper-limb exoskeleton. 97. Ocampo R, Tavakoli M, editors. Visual-haptic colocation in ro- Wearable Robotics: Challenges and Trends: Springer; 2017. p. botic rehabilitation exercises using a 2d augmented-reality dis- 181–5. play. 2019 International Symposium on Medical Robotics 81. Irastorza-Landa N, Sarasola-Sanz A, Shiman F, López-Larraz E, (ISMR): IEEE; 2019. Klein J, Valencia D, et al. EMG discrete classification towards a – myoelectric control of a robotic exoskeleton in motor rehabilita- 98. Gui K, Liu H, Zhang D. Toward multimodal human robot inter- action to enhance active participation of users in gait rehabilita- tion. Converging Clinical and Engineering Research on – Neurorehabilitation II: Springer; 2017. p. 159–63. tion. IEEE Trans Neural Syst Rehab Eng. 2017;25(11):2054 66. 82. Hogan N. Impedance control: an approach to manipulation: part 99. Jamwal PK, Hussain S, Ghayesh MH, Rogozina SV. Impedance I—theory. 1985 control of an intrinsically compliant parallel ankle rehabilitation – 83. Hogan N. Impedance control: an approach to manipulation: part robot. IEEE Trans Ind Electron. 2016;63(6):3638 47. II—implementation. 1985 100. Song A, Pan L, Xu G, Li H. Adaptive motion control of arm 84. Alqaudi B, Modares H, Ranatunga I, Tousif SM, Lewis FL, Popa rehabilitation robot based on impedance identification. Robotica. DO. Model reference adaptive impedance control for physical 2015;33(9):1795–812. human-robot interaction. Control Theory Technol. 2016;14(1): 101. Cabitza F, Locoro A, Banfi G. Machine learning in orthopedics: a 68–82. literature review. Front Bioeng Biotechnol. 2018;6:75. 85. Li Z, Huang Z, He W, Su C-Y. Adaptive impedance control for an 102.• Fong J, Ocampo R, Gross DP, Tavakoli M. Intelligent robotics upper limb robotic exoskeleton using biological signals. IEEE incorporating machine learning algorithms for improving func- Trans Ind Electron. 2016;64(2):1664–74. tional capacity evaluation and occupational rehabilitation. J 86. Figueiredo J, Félix P, Santos CP, Moreno JC, editors. Towards Occup Rehabil. 2020; This study shows how a machine learn- human-knee orthosis interaction based on adaptive impedance ing algorithm in conjungtion with a robotic terpist can work control through stiffness adjustment. 2017 International to evaluate and improve the therapeutic ionntervention. Conference on Rehabilitation Robotics (ICORR): IEEE; 2017. 103. Dolatabadi E, Taati B, Mihailidis A. An automated classification 87. Geoffroy P, Bordron O, Mansard N, Raison M, Stasse O, Bretl T, of pathological gait using unobtrusive sensing technology. IEEE editors. A two-stage suboptimal approximation for variable com- Trans Neural Syst Rehab Eng. 2017;25(12):2336–46. pliance and torque control. 2014 European Control Conference 104. Cui C, Bian G-B, Hou Z-G, Zhao J, Su G, Zhou H, et al. (ECC): IEEE; 2014. Simultaneous recognition and assessment of post-stroke 88. Azimi V. Model-based robust and adaptive control of transfemoral hemiparetic gait by fusing kinematic, kinetic, and electrophysio- prostheses: theory, simulation, and experiments: Georgia Institute logical data. IEEE Trans Neural Syst Rehab Eng. 2018;26(4): of Technology; 2020. 856–64. 89. Torabi M, Sharifi M, Vossoughi G. Robust adaptive sliding mode 105. Badesa FJ, Morales R, Garcia-Aracil N, Sabater JM, Casals A, admittance control of exoskeleton rehabilitation robots. Sci Iran Zollo L. Auto-adaptive robot-aided therapy using machine learn- – Trans B Mech Eng. 2018;25(5):2628 42. ing techniques. Comput Methods Prog Biomed. 2014;116(2): 90. Burgar CG, Lum PS, Shor PC, Van der Loos HM. Development 123–30. of robots for rehabilitation therapy: the Palo Alto VA/Stanford 106. Barzilay O, Wolf A. Adaptive rehabilitation games. J – experience. J Rehabil Res Dev. 2000;37(6):663 74. Electromyogr Kinesiol. 2013;23(1):182–9. 91. Serpelloni M, Tiboni M, Lancini M, Pasinetti S, Vertuan A, 107. Cai S, Li G, Su E, Wei X, Huang S, Ma K, et al. Real-time Gobbo M, editors. Preliminary study of a robotic rehabilitation detection of compensatory patterns in patients with stroke to re- system driven by EMG for hand mirroring. 2016 IEEE duce compensation during robotic rehabilitation therapy. IEEE J International Symposium on Medical Measurements and Biomed Health Inform. 2020. Applications (MeMeA): IEEE; 2016. 92. Sarasola-Sanz A, Irastorza-Landa N, López-Larraz E, Bibián C, Helmhold F, Broetz D, et al., editors. A hybrid brain-machine Publisher’sNoteSpringer Nature remains neutral with regard to jurisdic- interface based on EEG and EMG activity for the motor tional claims in published maps and institutional affiliations.