A prerequisite for both robot- and therapist-assisted rehabilitation is that individuals must maintain their ability to adapt to new dynamic environments (Shadmehr and Mussa-Ivaldi, 1994). Indeed, implicit motor adaptation may be able to reshape the altered sensorimotor mappings and contribute to cortical reorganization, potentially limiting the consequences of irreversible tissue damage in normal-appearing brain tissue and MS lesions (Reinkensmeyer et al., 2004; Rocca et al., 2005). For this reason, adaptive training protocols that introduce unfamiliar dynamic environments for individuals to adapt to, rather than simply assisting them during movement practice, may be beneficial to PwMS (Patton and Mussa-Ivaldi, 2004). Although PwMS have demonstrated residual capabilities for sensorimotor adaptation in arm and posture control, cerebellar deficits have been linked to difficulties in adapting to novel dynamic environments (Maschke et al., 2004; Smith and Shadmehr, 2005). It has been shown that individuals with cerebellar degeneration lose the motor learning mechanism based on the feed-forward control component and involved in motor adaptation (Maschke et al., 2004; Smith and Shadmehr, 2005). Individuals with MS still display this mechanism, albeit impaired (Leocani et al., 2007; Casadio et al., 2008) in such a way that could contribute to the coordination deficit and tremor associated with MS. Since force field adaptation exercises can train this feed-forward control mechanism, they may be effective in reducing tremor, improving upper limb coordination, and reducing disability in PwMS. Adaptive training may, therefore, be a promising rehabilitation approach for PwMS who exhibit various types and degrees of deficits. In the literature, this approach based on targeting sensorimotor adaptation in dynamic environments can be found in some protocols tested on PwMS (Carpinella et al., 2009; Vergaro et al., 2010; Basteris et al., 2011; Solaro et al., 2020). All these cited studies employed the Braccio di Ferro (Casadio et al., 2006) to develop an 8-session-long rehabilitative protocol involving robot-based reaching movements. In some of these studies (Carpinella et al., 2009; Basteris et al., 2011; Solaro et al., 2020), the task consisted of a series of reaching movements with a position-dependent resistive force directed along the line that connected the end-effector to the target, designed to challenge muscle weakness. However, the instability of the environment was given by additional force fields: a velocity-dependent force perpendicular to the instantaneous movement direction in Carpinella et al. (2009), and a virtual point mass connected to the subjects’ hand through a linear spring, that acted as “virtual tool”, in the remaining studies (Basteris et al., 2011; Solaro et al., 2020). Differently, in the protocol of Vergaro et al. (2010), by means of an iterative procedure, the robot learned the forces necessary to generate a perturbation directed orthogonally with respect to the trajectory that, for each target direction, either enhanced or decreased the lateral deviation of the average trajectories of the subject, estimated during a baseline session. Indeed, the procedure used in some works (Vergaro et al., 2010; Basteris et al., 2011; Solaro et al., 2020) was to calculate forces and spring stiffness at the beginning of each session and to let the protocol adapt its difficulty to the subject’s specific impairment and the improvements - if any - that occurred from session to session. The results of these works showed that PwMS revealed a preserved ability to adapt to robot-generated forces, greater in subjects with non-cerebellar symptoms (Solaro et al., 2020). In particular, subjects showed smoother and more linear movements (Vergaro et al., 2010; Solaro et al., 2020) over and within sessions. In addition, several studies have reported a significant improvement in the Nine-Hole Peg Test (9HPT) following training (Carpinella et al., 2009; Vergaro et al., 2010; Basteris et al., 2011; Solaro et al., 2020), indicating a potential transfer of therapy benefits to activities of daily living. Although the 9HPT primarily assesses manual dexterity, it requires coordination of the entire limb. Therefore, even though when using Braccio di Ferro the hand is not actively involved in the reaching exercise, the improvement observed in the 9HPT may be linked to enhanced coordination in the elbow and shoulder.

Robotic devices are capable of providing haptic feedback in a controlled manner, not only to perturb the environment but also to assist in movement execution. Given the wide variety of symptoms and their different severity in PwMS (Lublin et al., 2014), assistance should be tailored according to the motor skills of the individual (Casadio and Sanguineti, 2012; Gassert and Dietz, 2018). Assistance can take the form of gravity support, guidance through specific movement patterns, or help with movement completion. A concern when providing excessive assistance is the “Slacking” effect (Casadio and Sanguineti, 2012), which refers to a reduction in voluntary movement control caused by repetitive passive mobilization of the limbs. To avoid this, a potential solution is to implement the real-time tailoring of the assistance according to the individual’s needs and actual abilities. This “assistance-as-needed” approach seeks to reduce the risk of patients becoming overly reliant on robotic assistance, which could decrease their level of participation and hinder the potential for neuroplastic changes (Wolbrecht et al., 2008). Robot-based personalized assistance has been used with PwMS in several studies (Xydas and Louca, 2012; Groppo et al., 2017; Mannella et al., 2021) employing the Braccio di Ferro (Casadio et al., 2006), the Wristbot (Iandolo et al., 2019) and on an end-effector haptic device handled through a pen-like stylus comparable to the PHANTOM (Xydas and Louca, 2012). The study of Groppo et al. (2017) proposed a 23-session-long protocol to deal with the progressive worsening of motor functions in one PwMS. This multidisciplinary protocol involved traditional occupational therapy and a robot-based task, during which the subject performed center-out reaching movements. After 2 s from the movement onset, unless the subject was able to reach the target on their own, a minimally assistive force modulated according to the hand speed was generated by the robot. Groppo et al. (2017) found signs of improved motor control, given the significant increase in both the velocity and the smoothness of arm trajectories during robot-based reaching movements. Additionally, the fMRI revealed that multidisciplinary rehabilitation in MS seems to be clinically efficacious and to have a significant impact on brain functional reorganization in the short-term (Groppo et al., 2017). Mannella et al. (2021) trained the most affected limb of 7 PwMS in a 4-week robot-based program. The task was a continuous tracking of a figure targeting continuous movements in the flexion/extension and radial/ulnar deviation 2-dimensional space, in the presence of an assistive force. The force was implemented as a spring pulling the subject toward the target, and its rigidity was modulated within and between sessions, according to the performance of the subject in terms of accuracy in tracing the path. Actually, when the performance reached a specific level, assistance switched to resistance and pushed the subject far from the target, thus generating a dynamic environment to which to adapt. Similarly to what was found by Groppo et al. (2017), at the end of the treatment period, the authors detected a greater motor accuracy and control (Mannella et al., 2021), quantified by lower errors in tracking and tracing the target. In contrast, Xydas and Louca (Xydas and Louca, 2012) proposed an augmented version of the 9HPT, which incorporates assistive forces to transform it into a physiotherapy and rehabilitation system. The system included adaptive assistive forces based on healthy users’ reference target trajectories and was evaluated in a single session by three PwMS. The results showed a potential improvement in the upper limb performance in 3-dimensional reaching tasks, indicating that the system could be effectively used for rehabilitation in complex movements. Analogously to the personalization of the level of assistance, Octavia and Coninx (2014) adapted the difficulty level of the training tasks proposed. Based on the information about the training progress of the subject, the algorithm determined if and how the difficulty level should have been adapted. The study revealed that the participants followed different training patterns and progression, thus confirming the need for personalized levels of difficulty. This adaptive personalized training has shown to be beneficial and appreciated by users (Octavia and Coninx, 2014).

Robotics allows for repetitive and consistent motor movements at high dosages, however, repetition alone, without usefulness or meaning in terms of function, is not enough to produce increased motor cortical representations (Bayona et al., 2005). Rather than on the specific impairment, rehabilitation should focus on task-specific training to improve the performance in functional tasks through goal-directed practice and repetition. Task-specific training means practicing context-specific motor tasks while receiving some form of feedback (Schmidt and Lee, 1988). Robot-aided haptic feedback is a promising approach for rehabilitation, as it can provide information to supplement or substitute visual and auditory cues (Demain et al., 2013). This kind of feedback can help internalize the movements and increase proprioceptive awareness, thereby enhancing motor learning (Winter et al., 2022). Motor learning and skill acquisition are able to elicit the functional reorganization of cortical areas and the development of new motor pathways to restore limb function (Plautz et al., 2000). The Haptic Master has been used by a few studies (Octavia and Coninx, 2014; Feys et al., 2015; Maris et al., 2018) to train PwMS in an 8-session-long robot-based protocol additional to the conventional therapy. All exercises required accurate and stabilized end-positions to successfully perform the task-oriented movements. The exercises varied in the number of movement directions (1-2-3D), the haptic environment, the precision level and type of required movements, and the cognitive load. At the end of the training, movement tasks in three dimensions, measured with the robot, were performed in less time and more efficiently (Feys et al., 2015). Significant improvements were found in Maris et al. (2018) for shoulder ROM, handgrip strength, perceived strength, and Wolf Motor Function Test (WFMT) activities. Tedesco Triccas et al. (2022) investigated the impact of the intervention of Maris et al. (2018) on patients’ lives: 1) Participants felt that there was a positive impact of the training on strength, endurance, and during activities of daily living; 2) Participants expressed feelings of motivation and self-improvement about the system usage. Similarly, Gijbels et al. (2011) employed the Armeo Spring system to develop a mechanical-assisted therapy involving repetitive and active exertion of goal-directed movements, during the practice of complex motor tasks. The 8-week-long protocol, additional to conventional therapy, included the repetition of 5 tasks, ranging from gross movement, over more precise movement, to subtle strength-dosed movement. Significant gains were found in functional capacity tests [Upper extremity performance test for the elderly (TEMPA), 9HPT, Action Research Arm Test (ARAT)], particularly in subjects whose upper limb function was mostly affected at baseline. Gandolfi et al. (2018) compared two approaches for a 5-week-long training: a robot-assisted hand training using Amadeo that was mainly focused on visual feedback and had a task-specific approach, and a robot-unassisted training that dealt with functional movement and context-specific training. Both training protocols shared common features such as unilateral training, mobility, stretching, and exercise progression. However, only the robotic training involved more intensive, repetitive, and task-specific exercises. The main finding of this study was that upper limb activity and function improved after both treatments but only the robot-assisted hand training reported significant improvements in the assessment of skills in the life habits domain (Motor Activity Log (Taub et al., 1993)). In addition, preliminary observation of muscular activity showed enhancement of the extensor carpi radialis activation only in the robot-assisted group, suggesting a task-specific effect of this training mode on muscle activity. Finally, Feys et al. (2015) tried to investigate which types of robotic outcome measures could be clinically relevant. The protocol lasted 4 weeks and involved tasks that required motor accuracy, ROM, and the ability to exert high-speed movements, with different levels of assistance and difficulty. Significant correlations were found between specific functional measures (specifically ARAT and Purdue pegboard test) and movement tasks.

Robotic devices offer the opportunity to be synchronized with other electronic systems, in order to assess a broader set of body signals (Rizzoglio et al., 2020), deliver external additive feedback (Cuppone et al., 2016) and apply therapeutic stimulations (Sampson et al., 2016). The core instances of systems synchronized to robots are screen displays on which visual feedback of the virtual reality environment associated with the task is provided. Apart from these, it is hard to find examples of the use of devices in combination with robots for the rehabilitation of PwMS. The only study that apply was conducted by Sampson et al. (2016), combining the Armeo Spring (Sanchez et al., 2004) with Functional Electrical Stimulation (FES). The protocol was tested on 5 PwMS and involved 18 sessions. The robot provided kinematic data to a real-time processor that interfaced with custom FES hardware and the display. FES has been shown to be effective in augmenting strength in stroke (Glinsky et al., 2007) and spinal cord injury (Martin et al., 2012) and in reducing motor fatigue in MS (Chang et al., 2011). Through advanced model-based controllers, the passive robotic arm support combined with FES was meant to improve movement quality by promoting accuracy and voluntary effort and to increase the intensity of the intervention with minimal therapist input. Promisingly, the results showed improved accuracy of tracking performance both when assisted and unassisted by FES, a reduction in the amount of FES needed to assist tracking, and a decreased impairment in the arm trained.

In Section 1, we have presented the classification of MS in the various subtypes and their associated symptoms. Among the five most common symptoms there are weakness/numbness in one or more limbs and visual impairments, such as painful monocular visual loss or double vision (Rolak, 2003). Both these symptoms result in a crucial weakness of robotic rehabilitation in PwMS. The robotic devices employed for PwMS rehabilitation are designed for unilateral use: when both limbs are affected and need to be trained, they have to undergo the training subsequently and not simultaneously. This procedure needs a longer duration of each rehabilitative session and requires moving the device and repositioning the user. A possible, but expensive and not always feasible, solution would be to synchronize two robotic devices to allow bimanual tasks and reduce downtime (Albanese et al., 2023). On the other side, visual impairments play a crucial role since all the studies about rehabilitation in PwMS (Carpinella et al., 2009; Feys et al., 2009; 2015; Vergaro et al., 2010; Basteris et al., 2011; Gijbels et al., 2011; Xydas and Louca, 2012; Octavia and Coninx, 2014; Sampson et al., 2016; Groppo et al., 2017; Maris et al., 2018; Solaro et al., 2020; Mannella et al., 2021; Tedesco Triccas et al., 2022) employed visual feedback as the main informative feedback to users. Indeed, these studies involved a virtual environment displayed on a screen and all of them required the absence of visual deficit as an eligibility criterion to participate in the study. The issue here is that visual impairments could affect subjects’ performance, impacting both their motor skills and their motivation. Despite being a weakness, robots allow adding other feedback, such as auditive or haptic, to replace or be integrated with the visual one to allow conducting the training. Finally, we need to take into account that 40%–70% of PwMS reveal cognitive symptoms, such as memory loss and dementia (DeLuca et al., 2015). As for visual impairments, most rehabilitative protocols are not meant to deal with cognitive deficits and require a minimum score in cognitive assessment tests as inclusion criteria. Among these tests, a Montreal Cognitive Assessment (MoCA) (Ng et al., 2015) score > 15 was required by Manuli et al. (2020b) and a Mini-Mental (Pfeiffer, 1975) > 24 by other studies (Carpinella et al., 2009; Gijbels et al., 2011; Gandolfi et al., 2018; Solaro et al., 2020).

Despite the potential benefits of robotic devices in the rehabilitation of neurological conditions (neurorehabilitation), their application is limited by difficulties in the integration between patient and machine, as well as by the variability of clinical conditions among patients, which may contribute to these difficulties. Guidelines offered by both Resquín et al. (2016) and Huang et al. (2017) emphasize the importance of careful patient selection criteria for robotic rehabilitation. For instance, among stroke patients, they suggest admitting only patients with moderate motor skills based on the Fugl-Meyer and Motor Assessment Scale scores. Previous studies indicate that robotic training is more beneficial for individuals with moderate to severe deficits, while those with better motor function do not experience greater benefits from innovative device training compared to conventional training (Duncan et al., 1983; Kim et al., 2010; Morone et al., 2011). Given this and the high variability in MS severity, the inclusion criteria employed by the studies of robotic rehabilitation for PwMS have requirements concerning minimal motor skills and minimal levels of disability. These criteria involved the Motricity Index (Collin and Wade, 1990) between 50 and 84 (Gijbels et al., 2011) or between 14 and 25 considering only the performance of the shoulder (Maris et al., 2018), the Expanded Disability Status Scale (EDSS) (Kurtzke, 1983) < 7.5 (Carpinella et al., 2009; Manuli et al., 2020b; Solaro et al., 2020) or between 1.5 and 8 (Gandolfi et al., 2018), the 9HPT (Kellor et al., 1971) score between 30 and 180 (Carpinella et al., 2009; Solaro et al., 2020) or 300 (Gandolfi et al., 2018) seconds.

There are several general considerations that need to be taken into account about the mechanical design of the devices used for robotic rehabilitation in PwMS. For instance, to simulate realistic daily actions, the devices must provide an adequate number of DoFs to allow for proper joint rotations. However, it is natural that some of the devices used in rehabilitation have limitations in this regard. For example, the Braccio di Ferro allows for elbow flexion/extension and shoulder internal/external rotations but only permits planar movement (Casadio et al., 2006). The Amadeo focuses on finger flexion/extension without allowing abduction and adduction (Gandolfi et al., 2018), while the Armeo Spring enables only forearm pronation and supination and is not intended to enable other wrist movements (Gijbels et al., 2011). Furthermore, although most of the devices presented in this context target the upper limb, they are designed to mobilize and measure only some of the upper limb joints. The Amadeo and Wristbot devices measure finger movements and wrist rotations, respectively, while other joints are not assessed and are held in place (Iandolo et al., 2019). The Braccio di Ferro and the Armeo Spring measure and control the shoulder and elbow joints, but the movement of the wrist (except for forearm pronation/supination in the Armeo Spring) is constrained. In contrast, the Haptic Master and the PHANTOM allow all the upper-limb joints to move freely, and the position of the end-effector determines their configuration and vice versa Feys et al. (2009, 2015), in the same way as it happens during usual interactions with objects. In addition to these constraints, both the affordability and accessibility of robotic devices should be considered in this analysis. While robotic-based rehabilitation has been shown to be effective, the cost and limited availability of these devices remain an issue. Despite being commercially available, robotic devices are not mass-produced, and their cost remains high (Laut et al., 2016). Considering PwMS frequent difficulties in walking and moving independently, it would be an added value to make robot-based rehabilitation more accessible, by setting such systems in undersupervised environments outside of treatment centers.

Robot-based MS rehabilitation will be definitely interested in the technological advancements that are currently involving the field of robotics. Rapid advances in digital electronics, hardware speed and accuracy, and fabrication techniques (Bezzo et al., 2015; Gopura et al., 2016) have played a crucial role in lowering the expenses associated with the production of both research prototypes and commercial products. The emergence of real-time control software, combined with the significant reduction in computing costs, has led to the development of a plethora of sophisticated and accurately controlled robotic devices for rehabilitation (Nizamis et al., 2021). Furthermore, advances in software, signal processing, and machine learning are expected to continue offering incremental contributions to technologies and algorithms for neurorehabilitation at an ever-increasing pace (Nizamis et al., 2021). The use of large amounts of data to implement machine-learning approaches could be facilitated by improving the overall networking between devices. Overall, the ongoing multidisciplinary approach is currently encouraging further synergies between traditional research in physics and engineering, chemical, biological, and medical science to develop new applications and acquire new capabilities.

The goal of digital health is to improve healthcare outcomes, increase efficiency, and reduce healthcare costs by leveraging technology to optimize healthcare delivery and patient care. It encompasses a wide range of applications, including telemedicine, electronic health records, health data analytics, genomics, artificial intelligence, and mobile health apps. These technologies are being applied in various aspects of medicine, such as diagnosis, treatment, clinical decision support, care management, and care delivery (Mathews et al., 2019). Since the great promise held for improving healthcare outcomes, increasing efficiency, and reducing healthcare costs, the digital health sector has seen significant investment (Health, 2018). Digital health technologies provide a wealth of valuable data and connectivity, significantly amplifying the potential of robotics within healthcare. These technologies excel in the collection and storage of extensive patient data. By utilizing these vast datasets, it becomes feasible to identify critical trends and patterns, thereby strengthening research and the development of treatment strategies (Chang, 2023). Robotic systems can leverage this data to deliver personalized care, make real-time treatment adjustments, and facilitate monitoring, both in on-site and remote scenarios (Zhu et al., 2007; Barzilay and Wolf, 2013; Gross et al., 2013; Shirzad and Van der Loos, 2013). Advances in machine learning and deep learning have led to disruptive innovations in radiology, pathology, genomics, and other fields (Ramesh et al., 2004; MacEachern and Forkert, 2021). However, modern machine learning models require millions of parameters that need to be learned from sufficiently large, curated datasets to achieve clinical-grade accuracy while being safe, fair, equitable, and generalizing well to unseen data (Althnian et al., 2021; Varoquaux and Cheplygina, 2022). The issue here is how to address the problem of data governance and privacy by training algorithms collaboratively without exchanging the data itself. Rieke et al. (2020) proposed the approach of federated learning, which enables multiple parties to train collaboratively without the need to exchange or centralize data sets. Indeed, the machine learning processing occurs locally at each participating institution and only model characteristics are transferred. Concerning rehabilitation for PwMS, this approach has the potential to drastically increase sample size and enable a broader knowledge and an autonomous approach to the most correct practice to deal with the wide variety of symptoms of PwMS.

In the last decades, the demand for rehabilitation services has increased due to the aging phenomenon and the prevalence of chronic diseases, but it has faced a shortage of rehabilitation professionals and other barriers, such as low incomes, that deny access to rehabilitation services (Martinez-Martin and Cazorla, 2019). However, the continuity of exercises is crucial for the success of physical therapy and rehabilitation in PwMS, and the recommendation is to continue treatment practices in a home environment to support and maintain functional development in PwMS. Actually, at-home rehabilitation is suitable both for those requiring minor assistance and for those who cannot commute regularly to the treatment in physical therapy centers (Mayetin and Kucuk, 2022). Home-based rehabilitation has been found to be an effective alternative to traditional rehabilitation programs for stroke patients: it can be personalized and adjusted to the patient’s needs, can provide a more convenient and cost-effective therapy option (Catalan et al., 2018), achievable in a comfortable setting (Akbari et al., 2021). Additionally, unlike traditional exercises, where accountability relies on self-reporting, robotics enables objective monitoring, ensuring continuous exercise and program adherence. The COVID-19 pandemic caused an early discharge of existing patients, the suspension of new patients’ admissions, and the reduction of activities to decrease contacts (Manjunatha et al., 2021). Thus, the easiness of use and the low maintenance needed by haptic devices, together with the emerged necessity for further deploying the power of the internet for the purpose of communication and big data analysis, have attracted a lot of attention for home-based rehabilitation (Akbari et al., 2021). In particular, the need for effective home rehabilitation has received a boost from the need for early rehabilitative treatment to reduce long-term consequences in post-stroke patients (Akbari et al., 2021). Additionally, although many rehabilitation devices including robots are not available to be used at home due to their high costs, recent research has focused on low-cost devices able to assure both safety and effective results, in turn of a simpler design and functioning (Dowling et al., 2014; Rudd et al., 2019; Lambelet et al., 2020). The rehabilitation program for PwMS typically involves several sessions spread over a period of weeks or months. However, individuals with motor disabilities and limited mobility may face difficulties and expenses in traveling to health centers for treatment, which can pose a challenge in accessing rehabilitation services (Zasadzka et al., 2021). Even if this situation has been further complicated by COVID-19, the scenario might evolve into an opportunity for PwMS, as it happened for post-stroke patients. The future perspective is for an increase in the research interest in low-cost and easy-to-use robotic devices and in robot-based protocols for home-based rehabilitation in PwMS.

Spasticity is a common symptom in PwMS, characterized by an increase in the resistance of the joint to movement. This increased resistance is dependent on the velocity and caused by an amplification of the stretch reflex during the passive stretching of the joints (Trompetto et al., 2014). Therapeutic exercise treatments for individuals with high levels of spasticity can be challenging for therapists due to the patients’ considerable stiffness (Bohannon and Smith, 1987). This can create a barrier to the provision of sufficient rehabilitation therapy (Lee et al., 2017). Although most studies on robot-based rehabilitation in PwMS exclude individuals with spasticity (Carpinella et al., 2009; Sampson et al., 2016; Manuli et al., 2020b; Solaro et al., 2020), robots should be capable of addressing spasticity in a safe and appropriate manner. The proposed methods involve setting an appropriate torque range to manage spasticity-induced resistance or pausing movement when the resistance surpasses the motor’s threshold (Nam et al., 2017). The spastic muscle remains active for a certain amount of time with exponential decay of the resistance torque at the end of the range of motion where the robot is actuated against spasticity. This phenomenon allows the robot to be applied to a spastic limb even with a low-torque output motor without causing excessive loading. However, more precise sensing and control systems are necessary to deal safely with this symptom during PwMS rehabilitation.

Robot-based rehabilitation is not meant to replace traditional therapies, but rather to complement them (Li et al., 2021). Robotic systems are ideal for providing intensive, task-oriented motor training for patients’ limbs under the supervision of a therapist, and are part of a suite of rehabilitation tools that includes nonrobotic methods (Harwin et al., 2006). However, the need to use healthcare resources efficiently is increasing, and there are proposals to enhance the cost-effectiveness of rehabilitation programs, including hospitals, care homes, and home-based rehabilitation (Ward et al., 2008). While robot-assisted therapy has typically been supervised by trained personnel, the trend towards home-based and autonomous rehabilitation is driven by the need to reduce staff workload, despite concerns about the deskilling of therapists and doctors (Li et al., 2021). By allowing patients to use robotics unsupervised or semi-supervised, therapists can handle multiple patients simultaneously during robot-assisted therapy sessions or allow them to receive robot-assisted training at home to increase therapy dose and regularity (Rosati et al., 2007; Ranzani et al., 2021). However, safety is a critical issue for minimally-supervised systems as they should be operated without supervision, restricting the role of the rehabilitation team to planning and remote monitoring (Rosati, 2010). Safety requirements necessitate that the robot does not move patients beyond their range of motion, avoids pressure points on fragile skin, and is easy to clean and compliant with infection control policies (Li et al., 2021). Despite concerns about safety and efficacy in the absence of qualified staff, recent research by Ranzani et al. (2021) has shown that a powered robot-assisted therapy device can be safely and intuitively used with minimal supervision by chronic stroke patients, while still meeting usability and perceived workload requirements. Interestingly, the study found that usability was inversely related to age but not to the level of impairment, with the oldest subjects experiencing the worst usability results (Ranzani et al., 2021). Given the high-dose therapy needed by PwMS, this population would definitely benefit from the increased dose and the continuum of care coming from home-based and autonomous rehabilitation, which could progressively increase patients’ involvement and autonomy from the clinic to home. While the potential for unsupervised therapies to complement conventional therapies in real-world settings is significant, safety and efficacy issues still need to be addressed to employ robotic devices with minimal supervision.