On the enhancement of EMG-driven neuromuscular models for the runtime control of powered orthosis

The availability of accurate and comprehensive models of human limbs, combining high reliability and real-time operation, is required to develop seamless and intuitive human-machine interfaces. Biomechanist have developed complex models of the human lower limb, combining kinematics and kinetics data with neural signals for the purpose of studying human motor control strategies. The complexity prevent their application to situations with stringent real-time requirements. We are currently working on the enhancement of an EMG-driven model of the human lower extremity to achieve comprehensiveness, accuracy, and fast runtime execution. Starting from the very complex model developed by Lloyd et al. we have evaluated how this model can be enhanced to achieve higher performances in terms of computation time with no loss of prediction accuracy. The enhanced model will be applied to the control of powered orthosis and to the development of advanced biomimetic control systems for humanoids robots. We started our investigation with the analysis of the impact of modeling the tendon as an infinitely stiff body and quantitatively evaluated the changes in the behavior of the modified model w.r.t. the original one. We also integrated a runtime anatomical model that allowed to execute the whole EMG-driven model at runtime. This is a significant improvement as the previous available model could not entirely be executed at runtime due to the complexity of the original anatomical model.

Experiments were performed at the Gait Laboratory of the School of Sport Science Exercise and Health of the University of Western Australia. A 12 camera motion capture system (Vicon, Oxford, UK) was used to record dynamic movements. EMG signals were collected from 13 selected muscles using double-differential surface electrodes with a sampling frequency of 2Khz. Ground reaction forces were recorded from 2 force plates with a sampling frequency of 100Hz. A dynamometer (Biodex, New York, USA) was used to perform calibration trials under isometric and isokinetics conditions. Tests involved ground level walking at different paces. A group of 6 male subjects was recruited. For each subject a minimum of 10 gait trials were recorded. The dataset comprised a total of 75 trials.

We evaluated the impact of the introduced changes on both the prediction accuracy and the computation performance. The flexion-extension knee torque computed by the enhanced model was compared to the torque estimated by the original model previously developed. Furthermore, estimated torques were compared to the reference torque recorded by the motion capture system. The stiff tendon model generated significantly better estimated w.r.t. the original model proving that the assumption on tendon stiffness did not compromise the prediction capabilities of the system. Model computation performances were tested on an embedded system that could potentially be used for the control of a powered orthosis. Results prove that the whole system is executed in just a fraction of the trial duration time, resulting in a model suitable for situations with real-time constraints. The enhanced model executed 100 times faster than the original one on average per subject. The experimental results support our hypothesis that high reliability in prediction of joint torques can be achieved from neural excitation within the imposed time limits.