Goal-oriented, subject adaptive, robot-assisted locomotor learning (GOALL)

Propulsion is a fundamental component of walking, and our lab is working to develop methods to train propulsion in individuals with neuromotor impairment.

So far, our group has pursued two methods for training propulsion during walking. The first method is based on belt accelerations. Our method uses a split belt treadmill and trains propulsion by accelerating the treadmill belt of the trailing limb during push off. Belt accelerations require subjects to produce greater propulsive force to maintain their position on the treadmill and increase trailing limb angle through increased velocity of the accelerated limb. In our first study with this protocol, we established that exposure to belt accelerations induced after effects in propulsion mechanics, as shown by increased self-selected speed in a training group, which was not observed in a velocity-control group.

Treadmill belt accelerations during intervention as a function of gait cycle


Experimental Protocol Schematic. Highlighted phases at the bottom signify periods in which kinematic marker data were collected. In all Baseline and Post- Training conditions, a user-driven treadmill controller allowed participants to walk at a self-selected speed. Each training condition included an initial one minute long ramp phase to gradually introduce belt accelerations or change in velocity. Top: Acceleration training. Belt accelerations are shown on the y- axis, where ε signifies the magnitude of accelerations applied during training (2 m/s2 or 7 m/s2). Bottom: Velocity Control training. Belt velocity is shown on the y-axis. During training an increase of ∆V = 0.05 m/s was applied over the final velocity achieved in Baseline.


Within participant change in gait speed from baseline walking, broken down by group (HA: High accelerations (7 m/s^2), LA: Low accelerations (2 m/s^2), VC: Velocity Control). Top Left: Group average change in GS across the experimental protocol, resampled in time. Shaded areas depicts standard error. Estimated change in gait speed is reported for acceleration groups based on duration of applied accelerations. Top Right: Mean and standard error of group average change in GS in experimental phases of interest ( BL: baseline, E. TR: Early Training, L TR: Late Training, E. PT: early post-training, L. PT: late post-training).. Asterisks denote significant change from baseline from post-hoc Tukey HSD analysis. Bottom: Histogram of responder analysis of after effects measured in the Early (right) and Late (left) post-training session.

Our second method is based on the application of pulses of joint torque to the hip and knee joint using a robotic exoskeleton. We first conducted a biomechanical analysis to establish how humans modulate their joint torque under a factorial combination of gait speed and stride length, two parameters that are crucial for propulsion mechanics. We conducted an experiment with healthy control subjects instructed to walk on a treadmill at various speeds and asked to modulate stride length via visual feedback.  Sagittal plane joint torques were extracted via an inverse dynamics analysis of instrumented treadmill and motion capture data.

We utilized a torque pulse approximation analysis to determine optimal timing and amplitude of torque pulses that approximate the difference in joint torque profiles resulting from different stride length conditions measured at different values of walking speed.  Our group analysis generated a set of 16 pulse torque assistance profiles that were experimentally tested using our ALEX exoskeleton, with 2 active d.o.f., actuating about the hip and knee joints. In our first experiment, healthy control participants were exposed to the 16 joint torque profiles in single strides. In a second experiment, healthy control participants were exposed to a selected subset of 8 joint torque profiles in repeated strides.  The effects and after-effects on hip extension and normalized propulsive impulse were assessed. 

We observed that on the group level, participants showed adaptation to late flexion torque and early extension torque and learning effects with late extension torque and early flexion torque – in which all after-effects were positive.  For the measure of normalized propulsive impulse we observed adaptation effects in which early pulse application for both late flexion torque conditions exhibited positive modulation and the early flexion torque condition exhibited significant positive modulation in late after-effects. 


We are interested in incorporating the subject responses in human-in-the-loop optimization methods to directly target propulsion mechanics during training. Instrumental to this goal, we have established whether Gaussian process modeling can be used to describe data collected from our previous experiment. Using subject response of 16 pulse torque assistance profiles and optimal length scale hyperparameters for each input parameters, Gaussian process model of group-level hip extension and propulsive impulse data are fitted and shown in above figures. Red dashed line: model at the stride before intervention (stride −1); blue line: model at the stride of intervention (stride 0). Our analysis shows that the effect of pulses of torque on propulsion mechanics can be described using Gaussian process modeling, which opens the possibility for stochastic optimization methods such as Bayesian Optimization that we plan to test in coming experiments. (G. Kim and F. Sergi pre-print)

Publications on this topic

A. Farrens, M. Lilley, F. Sergi, “Training Propulsion via Acceleration of the Trailing Limb”, IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 28, no. 12, 2020 – doi: 10.1109/TNSRE.2020.3032094, pre-print, available online.

A. J. Farrens, R. Marbaker, M. Liley, F. Sergi, “Training propulsion during walking: adaptation to accelerations of the trailing limb”, 16th International Conference on Rehabilitation Robotics, pre-print, available online.

R. L. McGrath, M. Pires-Fernandes, B. Knarr, J. S. Higginson, F. Sergi, “Toward goal-oriented robotic gait training: the effect of gait speed and stride length on lower extremity joint torques”, IEEE/RAS-EMBS International Conference on Rehabilitation Robotics, London, UK, August 2017, available online, pre-print.

R. L. McGrath, M. L. Ziegler, M. Pires-Fernandes, B. A. Knarr, J. S. Higginson, and F. Sergi, “The effect of stride length on lower extremity joint kinetics at various gait speeds,” PLoS One, vol. 14, no. 2, p. e0200862, Feb. 2019 – doi: 10.1371/journal.pone.0200862, pre-print, available online.

R. L. McGrath, F. Sergi, “Single stride exposure to pulsed torque assistance provided by a robotic exoskeleton at the hip and knee joints”, 16th International Conference on Rehabilitation Robotics, pre-print, available online.

R. L. McGrath, B. Bodt, F. Sergi, “Robot-aided Training of Propulsion During Walking Effects of Torque Pulses Applied to the Hip and Knee Joints During Stance”, IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 28, no. 12, 2020 – doi: 10.1109/TNSRE.2020.3039962, pre-print, available online.

Print Friendly, PDF & Email