use of robotic devices for objective and quantitative assessment of neurorecovery, not just the delivery of therapy. In the past 10 years, researchers have increasingly focused on the design of rehabilitation robots for the hand and wrist, because the ability to self-feed, groom, and care necessitates recovery of hand function and dexterity. In contrast to the periodic nature of walking, upper limb and hand movements involve dozens of degrees of freedom, leading to complex kinematic designs of rigid arm and hand exoskeletons and tendon or cable-based actuation schemes that attempt to reduce device weight and inertia by remotely locating the actuators (34). Some groups have embraced soft robotic technologies for glove-based designs that focus on functional grasps, using pneumatic actuation that may even facilitate home-based rehabilitation (35). There have been impressive advances in control methods for rehabilitation robots in the past decade, predominantly those that facilitate cooperation between robot and patient. Increasingly advanced methods to estimate the capability of the patient to initiate or execute reaching movements or gait trajectories have been proposed, which are coupled to adaptive control schemes for the robotic device to automatically adjust the amount of robot support on the fly, maximizing the patient’s contribution to movement execution [see (36) for an example in upper limb rehabilitation and (37) for lower limb rehabilitation]. This strategy is known to promote neuroplasticity, which is critical to recovery of movement coordination (38). Patient engagement, both cognitive and physical, is another factor known to promote neuroplasticity during rehabilitation (39). In the past decade, researchers have developed new methods to detect movement intent from patients using surface EMG to measure electrical activity of the muscles themselves or electroencephalography (EEG) to infer intent from changes in the electrical Downloaded from https://www.science.org on November 16, 2021 Dupont et al., Sci. Robot. 6, eabi8017 (2021) 10 November 2021 SCIENCE ROBOTICS | REVIEW 8 of 15 potentials recorded from the surface of the scalp. Clinical evaluation of these techniques is in the early stages, although some initial findings show that outcomes for EEG-based intent detection are comparable with robotic therapy without intent detection [see, for example,(40)]. Although this may at first seem to be a disappointing outcome, the number of movement repetitions achieved in a single therapy session using this technology is substantially lower than robotic rehabilitation alone given the complexity of the experimental setup and computational overhead. Despite this, the clinical gains are comparable, meaning that such technologies may enable more severely impaired individuals who cannot initiate movement to benefit from robotic rehabilitation. A final area of advancement in the past decade is in the application of robotic rehabilitation devices as assessment tools. Clinical assessment scales are known to be relatively coarse in their ability to detect improvements in motor function. Robotic devices, outfitted with high-resolution sensors, can be used to assess range of motion, intra- and interlimb coordination, and movement smoothness, among other features (41). In addition, these devices can track recovery over higher-resolution time scales, because data can be collected at each treatment session. There is great potential for robotic assessment of neurorecovery to influence the intervention itself, which gives promise to the potential for robotic devices to appreciably improve rehabilitation outcomes in the future. The developments of the past decade are starting to be assessed clinically, using both research grade devices and those that have been commercialized. Clinical studies aimed at the evaluation of efficacy of novel devices, controllers, and methods for detecting user intent for stroke and spinal cord injury rehabilitation are, in some cases, actively recruiting participants, whereas other studies are listed in the clinical trials database but are not yet recruiting. Example clinical studies include investigations of soft robotic gloves, interactive exoskeletons for gait rehabilitation, and the potential for using EMG or EEG to control a rehabilitation exoskeleton. Although not directly related to advances in robotics, there are additional clinical studies that aim to determine the efficacy of existing devices for treatment of different neurological impairments. For example, devices developed to treat stroke populations are being evaluated on spinal cord injury populations. Another notable ongoing effort is the evaluation of the efficacy of combining robotic rehabilitation with other therapeutic interventions, such as spinal stimulation or pharmacological treatments. Although robotic devices have been shown to effectively deliver therapy to both the upper and lower limbs after stroke and spinal cord injury, the improvements in clinical outcome measures of function to date have been modest when compared with traditional therapy (38). Future research efforts are increasingly focused on gaining a better understanding of the mechanisms of neuroplasticity, including how it can be reliably induced and exploited to maximize therapeutic outcomes. Such efforts are increasingly dependent on advances