Severe spinal cord injury (SCI) permanently abolishes motor functions caudal to the lesion. However, the neuronal machinery sufficient to produce standing and stepping is located below most SCI, and can be reactivated with electrical spinal cord stimulation, pharmacological cocktails, and task-specific training. In combination, these interventions can promote the recovery of coordinated stepping with full weight bearing capacities in the total absence of supraspinal influences. This impressive recovery of function relies on the ability of spinal circuitries to utilize multisensory information as a continuing source of motor control and as a substrate for learning after the loss of brain input. I will show that spinal circuits can recognize task-specific sensory input and instantly modulate or transform the patterns of muscle activity in order to execute a variety of motor tasks ranging from standing to walking, running, stepping backward or even jogging in the sideward direction. In turn, repetitive exposure to specific sensory patterns with practice allows for the significant optimization of these sensorimotor processes whereby spinal circuitries can learn to produce specific motor states in the total absence of brain input. Throughout my talk, I will emphasize the implication of the sensory control of locomotion and sensory-dependent remodeling of spinal circuitries with training for the design of useful rehabilitative robot.

Existing gait rehabilitation robots, such as the Lokomat, the Lopes, the PamPogo, the ALEX or the GaitTrainer facilitate treadmill training. Not all of them do however incorporate current state-of-the-art knowledge in motor learning and neuroscience. I will provide an overview over current gait robots and summarize the neuroscientific findings provided by the other speakers. I will transfer these findings to the current gait robots and will derive design guidelines that will enable science-driven development of new rehabilitation devices. Hardware- and software-related issues will be covered, and existing robots will be evaluated according to these guidelines.

Over the last decade task oriented locomotor training regimes like body weight supported treadmill training have been established in clinics as effective tools to improve the gait capacity of i.e. incomplete spinal cord injured subjects. The basis of this therapy is formed by the enhancement of neuroplasticity on a spinal and supraspinal level. This is achieved by providing a physiological multisensory input necessary to generate and shape the locomotor pattern of the spinal locomotion circuitry. With robotic therapy devices the physical constraints of the manually assisted training can be overcome and longer training sessions under physiological conditions are possible. Most locomotion robots have mainly been inspired by industrial robots, where a precise and reproducible generation of the end effector trajectory is the main design criterion. However, recent findings suggest that a certain degree of variation around the reference trajectory is therapeutically relevant. Additionally the therapy should follow the principle of "supporting and challenging". Therefore, multiple sensor systems have to be integrated in modern locomotion robots for implementation of "assist-as-needed" control strategies.
The successful application of locomotion robots is not only determined by technical specifications, but also by user demands. The currently available machines can only be applied in a clinical setting due to their complexity, price and need for specifically trained operators. However, in Europe the initial length of stay in a rehabilitation center has tremendously decreased over the last years due to economical restrictions. A careful study of the literature suggests that a moderate training intensity applied over a period of several months is much more effective than a highly intensive training for some weeks with long session durations and several session repetitions per day. This results in a high need for devices that can be used in a home environment by the patients independently from any personal supervision by an experienced therapist. Transferring locomotion robots to the patients' homes brings up new challenges in terms of patients' safety, autonomous use, size and weight. For optimization of the latter answers to several open questions have to be find: Which are the key elements of a locomotion robot to be therapeutically effective? Which training protocols have to be applied that the training effects in a robotic device may be transferred to improvements in real world situations?As a fist step into the direction of home based locomotion robotics the basic principles of a novel gait rehabilitation robot inspired by biological principles will be presented together with the results of a clinical pilot study in incomplete SCI subjects.

All moving systems are subject to the laws of physics. In control engineering of humanoid locomotion, this fundamental constraint is recognized by developing controllers around dynamic gait models. However, it is often overlooked when investigating human motor control. In the first part, I will show how a fundamental approach from dynamics to motor control generates neuromuscular models of legged locomotion which not only produce human walking but also predict muscle activation patterns observed in major leg muscles. The main ingredients of these models are sensory feedback controls that embed core dynamic models of legged systems in the physiological structures of the human body. In return, these embedded feedbacks suggest new ways of controlling robotic legs, which has led to the first powered ankle prosthesis that propels amputees and seamlessly adapts to the environment.
While these initial successes support the fundamental approach, generalizing it on a large scale presents another challenge. (i) It remains to be seen if multi-segment robot legs can be controlled by the identified feedbacks to replace or restore more human leg functions than the ankle prosthesis does. (ii) The embedded control in the neuromuscular models needs consolidation to accommodate changes in gait, speed and direction. And (iii) the underlying dynamic gait models have to grow in dexterity to capture more complex human behavior. I will outline in the second part how we attack these problems to generalize the fundamental approach from dynamics to motor control to complex systems in humanoid and rehabilitation robotics.

Locomotor patterns must be continuously adjusted to deal with changes in the environment (e.g., ice), body (e.g., fatigue), and other conditions (e.g., high-heeled shoes). When these changes are sustained and predictable, individuals can modify their internal model to account for them via motor adaptation, which is a form of short timescale learning. Specifically, we define adaptation as the trial-and-error process of changing a movement pattern in response to predictable circumstances. An important feature of adaptation is that the new pattern is stored and expressed as an aftereffect when conditions return to normal. Here, we will focus on a form of adaptation that occurs when people walk on a split-belt treadmill, which has two belts that can drive each leg at a different speed. Practicing split-belt walking changes the coordination between the legs and the resultant aftereffects have been shown to improve abnormal coordination patterns in individuals with stroke and other forms of cerebral damage. We will discuss the principles underlying locomotor adaptation and how we can apply these in rehabilitation settings. We will also highlight current work aimed at designing devices that simulate the action of the split-belt treadmill and can be worn during natural walking over the ground.

Because humans are bipeds most studies on gait have focused on leg movements. Similarly, rehabilitation of gait is seen in terms of robotic assistance of leg movements with little attention given to arm movements. The question is whether this is appropriate. The current view is that, during gait, each limb is controlled by special spinal circuits, termed CPG's (central pattern generators) and that these CPG's are interconnected through long projecting propriospinal neurons. If so one would expect that arm movements can have a beneficial effects on leg movements. Some evidence has been obtained in this direction in work on SCI patients on a treadmill. However, facilitation by arm movements is difficult to prove during gait because swinging the arms affects the biomechanics of the legs. For this reason several investigators have used recumbent stepping instead since there is good evidence that the same locomotor circuits are involved as in gait. One such study failed to find facilitation of leg movements by arm motion but in that case the leg muscle contractions requested were maximal, leaving little room to study facilitation. The question was therefore re-examined using submaximal contractions. In healthy subjects a beneficial effect of arm movements on leg locomotor output was found. It is argued that the use of robotic devices for gait should allow accompanying arm swing.

The robot allows standardized training sessions that simultaneously provide objective measures about the physical aspects of the training performed (e.g. applied forces, velocity, duration of training, leg excursions) and about the training effects, (i.e. the progress of recovery can be monitored). For an effective training essential cues have to be provided by the robotic device. For example, load- and hip joint related input is important for an appropriate leg muscle activation during a locomotor training (Dietz et al., Brain, 2002). Thus body re-loading as much as possible during the course of training and extensive hip joint excursion movements are essential to achieving an appropriate leg muscle activation and, consequently, to strengthen training effects. Other cues have still to be elaborated such as the best type of feedback information for the patient during a functional training episode, and how should it best be delivered to reinforce training effects. Furthermore, how much and how early after a SCI or brain damage should a patient be challenged during the locomotor training? In which way can virtual reality (VR) interventions enhance the effectiveness of training? Do only certain patient subgroups, e.g. children or elderly patients profit from such an approach? Or can VR make the training just more attractive to the patient? All these questions will be discussed during the workshop.

All rehabilitation robots (i.e. including lower-limb exoskeletons) need to manage the human-robot interface. Any human movement could be either a perturbation (to be "rejected" by the robot), or a voluntary action (requiring an adaptation of the robot controller). EMG-based assistive robots neglect the former, while stiff rehabilitation robots neglect the later. Here, we propose a new approach to address HRI during walking assistance, based on adaptive oscillators - a sort of fundamental dynamic motor primitive for periodic movements. As such, our approach (i) is trajectory-free (the user can flexibly change the desired movement features), (ii) requires simple sensors (only the assisted joints position), and (iii) provides an intuitive human-robot interface.
I will briefly review our first experiment, which was designed as a simple proof-of-concept: Assistance of periodic elbow movements. Afterwards, I will discuss the extension of these results to the case of walking assistance. The extended adaptive oscillator is capable of predicting the angular position of the user's joints in the future, based on the pattern learned during preceding cycles. Assistance is then provided by attracting the joints to this future position using a force field in a compliant lower-limb exoskeleton. To conclude, I will discuss the relevance of designing assistance protocols based on adaptive oscillators (or primitives in general), paving the way to the design of new rehabilitation protocols.

Systems theory suggests that the use of kinematic and dynamic models allow for more versatile and robust control of arms, legs and bodies (e.g. compliant or impedance control). Biological experiments back up this systems theoretic conclusion. I will shortly review some of the evidence for model based control in Biology and the systems theoretical models that are supported by this evidence. I will then review our work on control and learning for dextrous robots that spans model-free to model-based approaches. First, I will show how implementing one of these model based controllers on a legged robot allows for actively compliant locomotion and improves robustness in face of uncertain and moving terrain without sacrificing generality and versatility of achievable movements. Model based controllers however have the drawback of having a dependence on a number of possibly unknown model parameters. Yet, as I will show the 'model-based world view' ultimately even allows for more robust and better performing model-free reinforcement learning algorithms. Our model-free reinforcement learning algorithm finds solutions that approximate a model based optimal variable impedance controller. To illustrate this finding, I will present examples of the application of our learning methods to grasping and manipulation.