Therapeutic and assistive exoskeletons and exosuits show promise in both clinical and real-world settings. Improving their autonomy can enhance usability, effectiveness, and cost efficiency. This Review presents a generic control framework for autonomous operation of upper and lower limb devices and reviews current advancements and future directions. We highlight how data-driven machine learning aids in intention recognition, synchronization, patient assessment, and task-agnostic control. In addition, we discuss how reinforcement learning optimizes control policies through digital human twins and how generative AI supports therapy planning and patient engagement. Richer patient-specific data and more accurate digital twins are needed for clinical validation and widespread deployment.

Until today, in the field of motor learning and rehabilitation, haptic controllers were mostly limited to teach simple tasks such as movements along straight lines, curves, or circles. However, commonly, real-life tasks consist of more complex movements such as in writing, rehabilitation, or surgery. In this paper, a novel haptic controller for robot-assisted learning is introduced. This hybrid path controller can cope with interfering path sections, while it also incorporates the common requirements of effective motor learning: it allows freedom for making spatial errors, free timing to explore the task dynamics, and adaptation to the current skill level of the user. In a practicability study with two different robots, results confirmed the full functionality of the controller and its applicability for a broad range of complex movements.

Rehabilitation robots allow for longer and more intensive locomotor training than that achieved by conventional therapies. Robot-assisted gait training also offers the possibility to provide objective haptic, visual, and auditory feedback to the patients and/or therapists within one training session and to monitor functional improvements over time. This chapter provides an overview of the technical approach for one of the most widely used systems known as “Lokomat” including features such as hip abduction/adduction actuation, cooperative control strategies, assessment tools, and augmented feedback. These special technical functions may be capable of further enhancing training quality, training intensity, and patient participation.

One key question in motor learning is how the complex tasks in daily life - those that require coordinated movements of multiple joints - should be trained. Often, complex tasks are directly taught as a whole, even though training of simple movement components before training the entire movement has been shown to be more effective for particularly complex tasks ('part-whole transfer paradigm'). The important implication of the part-whole transfer paradigm, e.g. on the field of rehabilitation robotics, is that training of most complex tasks could be simplified and, subsequently, devices used to train can become simpler and more affordable. In this way, robot-assisted rehabilitation could become more accessible. However, often the last step in the training process is forgotten: the recomposition of several simple movement components to a complete complex movement. Therefore, at least for the last training step, a complex rehabilitation device may be required.In a pilot study, we wanted to investigate if a complex robotic device (e.g. an exoskeleton robot with many degrees of freedom), such as the ARMin rehabilitation robot, is really beneficial for training the coordination between several simpler movement components or if training using visual feedback would lead to equal benefits. In a study, involving 16 healthy participants, who were instructed in a complex rugby motion, we could show first trends on the following two aspects: i) the part-whole transfer paradigm seems to hold true and therefore, simple robots might be used for training movement primitives. ii) Visual feedback does not seem to have the same potential, at least in healthy humans, to replace visuo-haptic guidance for movement recomposition of complex tasks. Therefore, complex rehabilitation robots seem to be beneficial for training complex real-life tasks.

Series Elastic Actuation decouples actuator inertia from the interaction ports and is thus advantageous for force-controlled devices. Parallel or even passive compliance can fulfill a complementary role by compensating for gravitational or periodic inertial forces or by providing passive guidance. Here, these concepts are combined in an underactuated six degree of freedom (DoF) compliant manipulator with one actuated DoF. The mechanism comprises a spring assembly in which each spring serves as an actuation element and simultaneously provides passive compliance in the unactuated DoF. The device is designed to assist weight shifting via controlled lateral forces on a human pelvis during treadmill walking and its eigenfrequencies are tuned to align with normal gait. Six-DoF force and torque sensing are realized via a model of the spring deformation characteristics in combination with low-cost inertial and optical sensors. Experimental evaluation demonstrates that the system can effectively follow physiological weight shifting with low interaction forces and also has little impact on remaining pelvis motions.

On the quest to bring function of prosthetic legs closer to their biological counterparts, intuitive interplay of their control with the user’s impedance modulation is key. We present two control features to enable more physiological and more user-adaptive control of prosthetic legs: a neuromusculoskeletal impedance model (NeurImp) including a reflexive component, and a human model reference adaptive controller (HuMRAC), which can be combined with the former. In stance-phase simulations, the NeurImp allowed to control a prosthetic leg with physiological knee joint angle and moment. When perturbations were applied, the HuMRAC reduced the resulting root mean square error (RMSE) between simulated and physiological reference angle by 96%. In a pilot experiment with two unimpaired and one amputee subject, gait with the NeurImp deviated more from a physiological reference than with a conventional visco-elastic impedance controller. Subjects, however, preferred the NeurImp. When adding the HuMRAC to either of the two impedance controllers, the RMSE between the actual and the physiological reference angle was reduced by up to 54%. Subjects confirmed this finding and reported an easier stance-toswing transition. Simulation and pilot experiment suggest that a reflex-based impedance controller combined with an adaptive controller may improve user-cooperative behavior of active knee exoprostheses.