Justin Fong

Learning control in robot-assisted rehabilitation of motor skills – a review

The key idea in iterative learning control is captured by the intuition of ‘practice makes perfect’. The underlying learning is based on a gradient descent algorithm iteratively optimising an appropriate input–output measured criterion. How this paradigm is used to model quantitatively, at an input/output level, the learning that happens in the context of human motor skill learning is discussed in this note. Experimental studies of human motor learning, in robotically controlled environments, indicate that a model consisting of a classical (iterative) learning control augmented with an appropriate kinematic model of human motor motion fits the observed human learning behaviour well. In the context of the rehabilitation of motor skills, such models promise better human–machine interfaces that extend the capability and capacity of rehabilitation clinicians by creating effective robot–patient–clinician feedback loops. The economic promise of robot-assisted rehabilitation is to greatly extend the intervention...

An investigation into the reliability of upper-limb robotic exoskeleton measurements for clinical evaluation in neurorehabilitation

Robotic exoskeletons are increasingly being used for neurorehabilitation, due to a number of perceived advantages. Once such advantage is the potential to use the large amounts of previously unavailable measurements to provide continuous assessment of the patient. This study investigates the validity of such measurements through an experimental protocol. Reaching movements within and outside an upper-arm rehabilitation exoskeleton (ArmeoPower) of 10 healthy subjects are compared using five commonly-used kinematic metrics (Peak Speed, Time to Peak Speed, Curvature, Smoothness, Accuracy). The study finds that (1) the robotic exoskeleton significantly affects the reaching movements of healthy subjects, (2) the measurements of the exoskeleton accurately represent the movements of the wrist, and (3) evolution of the in-exoskeleton movements over multiple sessions is indicative of changes in movements outside the robot, even though differences remain - suggesting that evolution of this data may be used to monitor patient progress.

Effects of robotic exoskeleton dynamics on joint recruitment in a neurorehabilitation context

Increasing research has been conducted into the use of robotic devices for neurorehabilitation. One advantage of these devices over traditional rehabilitation is the availability of measured data, which can be used to inform potential patient-specific protocol for recovery or simply to provide higher frequency feedback to the patients and therapists. It has previously been identified that such devices may have unplanned effects on the movement of patients. However, the exact nature of these effects are unknown, which makes the meaning of any measured data less clear. As such, this study investigates the effect of the mechanical dynamics of a robotic exoskeleton (ArmeoPower, Hocoma, Switzerland) on the movements of healthy subjects - particularly with respect to the movements of the shoulder, and joint utilisation. The study finds that the exoskeleton may encourage changes in shoulder movement in both magnitude and direction and changes in the joints recruited for the movement. Furthermore, the effects of the robot on joint utilisation are not consistent across reaching directions, however, the peak joint velocities are decreased across all joints and reaching directions.

A Tool to Address Movement Quality Outcomes of Post-stroke Patients

A new neuro-rehabilitation system is proposed to address the movement quality of post-stroke patients. The system is designed to be used concurrently with existing upper-extremity virtual rehabilitation devices, and to aide correction of compensatory trunk and shoulder movements. A 3D sensor is utilised to estimate the movement of the shoulder, and an auditory cue is given to the patient when the system estimates that a compensatory movement has been made. The results of preliminary trials of this system on a single patient are presented.

Dual-loop iterative optimal control for the finite horizon LQR problem with unknown dynamics

Abstract Achieving optimal performance over a finite-time horizon has gained a lot of attention in many engineering applications. Among them, the Finite Horizon Linear Quadratic Regulator (FHLQR) formulation for continuous-time linear time-varying systems has been well studied, with an optimal solution characterised by the Differential Riccati Equation (DRE). The solution of the DRE requires that the exact system dynamics are known. However, this assumption may not always hold, as the plant model might not be completely known or may change over time due to wear and tear. This paper proposes a dual-loop iterative algorithm to find the optimal solutions of the FHLQR formulation for continuous-time LTV systems. The inner loop utilises input trajectories based on an estimate of the optimal control gain with the addition of some excitation noise, and produces measured state trajectories. The outer loop improves the estimate of the optimal control gain utilising these measured state trajectories. It is shown in this work that with appropriate selection of the discretisation parameter T and the set of excitation signals, the proposed dual-loop iterative algorithm can converge to an arbitrarily small neighbourhood of the optimal solution. A simulation example demonstrates the effectiveness of the proposed method.

Modulation of shoulder muscle and joint function using a powered upper-limb exoskeleton

Robotic-assistive exoskeletons can enable frequent repetitive movements without the presence of a full-time therapist; however, human-machine interaction and the capacity of powered exoskeletons to attenuate shoulder muscle and joint loading is poorly understood. This study aimed to quantify shoulder muscle and joint force during assisted activities of daily living using a powered robotic upper limb exoskeleton (ArmeoPower, Hocoma). Six healthy male subjects performed abduction, flexion, horizontal flexion, reaching and nose touching activities. These tasks were repeated under two conditions: (i) the exoskeleton compensating only for its own weight, and (ii) the exoskeleton providing full upper limb gravity compensation (i.e., weightlessness). Muscle EMG, joint kinematics and joint torques were simultaneously recorded, and shoulder muscle and joint forces calculated using personalized musculoskeletal models of each subjects upper limb. The exoskeleton reduced peak joint torques, muscle forces and joint loading by up to 74.8% (0.113 Nm/kg), 88.8% (5.8%BW) and 68.4% (75.6%BW), respectively, with the degree of load attenuation strongly task dependent. The peak compressive, anterior and superior glenohumeral joint force during assisted nose touching was 36.4% (24.6%BW), 72.4% (13.1%BW) and 85.0% (17.2%BW) lower than that during unassisted nose touching, respectively. The present study showed that upper limb weight compensation using an assistive exoskeleton may increase glenohumeral joint stability, since deltoid muscle force, which is the primary contributor to superior glenohumeral joint shear, is attenuated; however, prominent exoskeleton interaction moments are required to position and control the upper limb in space, even under full gravity compensation conditions. The modeling framework and results may be useful in planning targeted upper limb robotic rehabilitation tasks.

Sensorless Force Estimator in Rehabilitation Robotics

Measuring the force exerted by patients in the exercise for rehabilitation after neurological injuries is important: in quantifying the patient’s motion capabilities, to ensure safety and to provide the appropriate amount of assistance, among others. Adding a force sensor for this purpose at the end-effector of a rehabilitation robot can add considerable cost. When a robotic device is dynamically transparent and mechanically backdrivable, a force estimator based on the model of the system can be used to estimate the force applied by the patient without using the explicit force sensor. This work validates the effectiveness of a model-based force estimator, derived from the literature, within the context of rehabilitation robotics, through a successful validation the strategy on the EMU upper-limb rehabilitation robot.

Calibration free upper limb joint motion estimation algorithm with wearable sensors

This paper aims to establish a post processing algorithm to estimate the upper limb motion, given a set of measurements from wearable sensors representing the orientation of the shoulder, upper arm and lower arm. The motivation of the development is the measurement of the upper limb motion for subjects with motor impairments, such as post-stroke patients preventing the use of specific motions for calibration purposes and allowing the sensors to be relatively insensitive to their mounting positions. The type of sensors has been left general, with the experimental validation in this paper carried out using inertial sensors and magnetic trackers. The method is validated both numerically and experimentally, and shows improvements compared to the common inverse kinematics approach, especially in the practical conditions where sensor mounting alignment is suboptimal.

Quantifying task similarity for skill generalisation in the context of human motor control

In this work, a simple model is used to characterize the learning behaviour of humans. Based on this model, it is possible to define a similarity measure between two tasks in order to quantify skill generalisation during the learning of simple motor tasks by humans. By fully exploring this similarity measure, a sequence of tasks capable of improving the learning efficiency for both healthy subjects and patients with motor impairment may be generated. A validation protocol is introduced and preliminary experimental results with six subjects are presented to validate the learning model and the similarity measure. Results show that the human learning of trajectory tracking tasks can accurately be modelled by an exponential decay of the average tracking error. The model fits well when the task is new or far away from a previously learnt task. Model parameters are used to analyse the learning performances of the subjects and the influence of previous tasks learning. Finally, it is shown that the similarity index can be constructed based on the proposed model to reflect skill generalisation.