Stable myoelectric control of a hand prosthesis using non-linear incremental learning

Abstract

Stable myoelectric control of hand prostheses remains an open problem. The only successful human-machine interface is surface electromyography, typically allowing control of a few degrees of freedom. Machine learning techniques may have the potential to remove these limitations, but their performance is thus far inadequate: myoelectric signals change over time under the influence of various factors, deteriorating control performance. It is therefore necessary, in the standard approach, to regularly retrain a new model from scratch. We hereby propose a non-linear incremental learning method in which occasional updates with a modest amount of novel training data allow continual adaptation to the changes in the signals. In particular, Incremental Ridge Regression and an approximation of the Gaussian Kernel known as Random Fourier Features are combined to predict finger forces from myoelectric signals, both finger-by-finger and grouped in grasping patterns. We show that the approach is effective and practically applicable to this problem by first analyzing its performance while predicting single-finger forces. Surface electromyography and finger forces were collected from 10 intact subjects during four sessions spread over two different days; the results of the analysis show that small incremental updates are indeed effective to maintain a stable level of performance. Subsequently, we employed the same method on-line to teleoperate a humanoid robotic arm equipped with a state-of-the-art commercial prosthetic hand. The subject could reliably grasp, carry and release everyday-life objects, enforcing stable grasping irrespective of the signal changes, hand/arm movements and wrist pronation and supination.

Keywords: force control; human–machine interfaces; incremental learning; machine learning; rehabilitation robotics; surface electromyography.

Figure 1

The setup of the first experiment. (A) Ten OttoBock MyoBock 13E200 sEMG electrodes, uniformly arranged (B) on the subject's forearm using a band of bio-compatible adhesive tape. (C) An ATI Mini45 force sensor, kept in place using double-sided tape.

Figure 2

(A) The i-LIMB Ultra pinch-gripping a soft foam ball. (B) The TORO humanoid robot, grasping a glass bottle.

Figure 3

nMSE and correlation coefficient per trial, in the batch setting: when training on the first three trials of the first session, and on the first three trials per session. Results for the linear RR and KRR methods are averaged over the ten subjects and the five DOFs. For the iRFFRR method, the results are in addition averaged over 25 runs with different random initializations.

Figure 4

Convergence of performance in terms of nMSE (A) and correlation (B) when increasing the number of RFFs D. The error bars indicate the intrasubject standard deviation over the 25 runs with different random initializations, averaged over the ten subjects and over the five DOFs.

Figure 5

nMSE and correlation coefficient per trial, in the incremental setting. For comparison, the results for KRR trained per session are also included. The presented results for the linear method are averaged over the ten subjects and the five DOFs. For the iRFFRR method, the results are in addition averaged over 25 runs with different random initializations.

Figure 6

nMSE and correlation coefficient per trial obtained by iRFFRR¹⁰⁰⁰, in the incremental setting (training on measured forces) and in the realistic setting 1 (training on graded stimulus) and 2 (training on binary stimulus). The results are averaged over 25 runs with different random initializations.

Figure 7

Typical measured and predicted forces by iRFFRR¹⁰⁰⁰ for the last trial of each session in the realistic setting 2 (training on the binary stimulus). Data taken from the first randomized run of the fifth subject.

Figure 8

nMSE and correlation coefficient per trial for individual DOFs obtained by iRFFRR¹⁰⁰⁰ in the realistic setting 2 (training on the binary stimulus). Results averaged over the 10 subjects and over 25 runs with different random initializations.

Figure 9

Comparison of the end-effector motion range (A), motion speed (B), and grasping force (C) of the end effector during each phase of Task 1. The mean values and one standard deviation are reported over the successful trials. (D) Motion speed during one typical trial. The color denotes the grasping force in arbitrary units, according to the color bar on the right-hand side.

Figure 10

Comparison of the wrist rotation range (A), wrist rotation speed (B), and grasping force (C) of the end effector during each phase of Task 2. The mean values and one standard deviation are reported over the successful trials. (D) Rotation speed during one typical trial. The color denotes the grasping force in arbitrary units, according to the color bar on the right-hand side.