Proceedings of the first workshop on Peripheral Machine Interfaces: going beyond traditional surface electromyography

Abstract

One of the hottest topics in rehabilitation robotics is that of proper control of prosthetic devices. Despite decades of research, the state of the art is dramatically behind the expectations. To shed light on this issue, in June, 2013 the first international workshop on Present and future of non-invasive peripheral nervous system (PNS)-Machine Interfaces (MI; PMI) was convened, hosted by the International Conference on Rehabilitation Robotics. The keyword PMI has been selected to denote human-machine interfaces targeted at the limb-deficient, mainly upper-limb amputees, dealing with signals gathered from the PNS in a non-invasive way, that is, from the surface of the residuum. The workshop was intended to provide an overview of the state of the art and future perspectives of such interfaces; this paper represents is a collection of opinions expressed by each and every researcher/group involved in it.

Keywords: EMG; human–machine interfaces; prosthetic control; prosthetics; rehabilitation robotics.

FIGURE 1

A myoelectric prosthesis implanted after shoulder disarticulation (A) 2-DOFs self-powered hand, wrist, and elbow, plus non-motorized mechanical shoulder with electrical fixation (B).

FIGURE 2

Closed-loop prosthetic control: including appropriate feedback for an increased embodiment.

FIGURE 3

Robustness, adaptability, and situational awareness (sensorimotor knowledge) as three complementary machine intelligence pursuits to enhance the expected clinical effectiveness of conventional and emerging myoelectric control systems.

FIGURE 4

An abstract representation of the use of situational awareness (knowledge) to supplement myoelectric control. In conventional myoelectric control, state information in the form of sEMG features is provided to the controller (A). Learned, prediction-based knowledge regarding the context (or contexts) of use can be used to modulate the parameters and the state-action mapping of a controller in a situation- and user-appropriate way (B).

FIGURE 5

Learned predictions can be used to adjust the control parameters of both conventional and emerging PMI controllers. Examples include using situational predictions to dynamically re-order control options, change controller gains, or adapt thresholds and filters such that they are matched to a user’s immediate needs.

FIGURE 6

Real-time machine learning provides up-to-date predictive state information to a control system. Temporally extended predictions can serve as supplementary state information to improve control performance, or may be directly mapped to a set or subset of the available control functions.

FIGURE 7

Conceptual design of the semi-autonomous control of prostheses. The basic idea is to enhance the artificial controller (processing unit) with an extra source of information (sensing interface) so that the system can operate automatically and autonomously, while the user has supervisory and corrective role. The main features of the system are automatic operation, bidirectional communication, semi-autonomous, and closed-loop control (see text for details). The flow of commands, sensor data, and feedback information are represented using blue, red, and green lines.

FIGURE 8

Example operation of a prototype system implementing semi-autonomous control of grasping in a dexterous prosthetic hand. The user wears augmented reality glasses equipped with a stereo camera pair and a stereoscopic “see-through” display. From top to bottom, the snapshots depict: (1) object targeting phase with augmented reality (AR) feedback about object selection, (2) automatic hand preshaping phase with AR feedback on the selected grasp type and aperture size, and (3) object manipulation phase. The panels on the right depict what the user actually sees through the glasses.

FIGURE 9

The Pisa/IIT SoftHand and the forearm adapter used to test the device on control subjects.

FIGURE 10

Average interaction torques (in mNm units) by controller type (Ajoudani et al., 2014).

FIGURE 11

General models of myoelectric interface interaction. (A) Interfaces with trained decoders. A decoder is trained to map sEMG signals (m) to human arm motion (y). Once trained, the decoder is used in real-time to estimate arm motion (y′) and map it to output (z) for an interface. (B) Interfaces utilizing motor learning. The brain adjusts the neural commands based on the interface output (z) by learning the inverse model of the decoder.

FIGURE 12

Embedded brain control for myoelectric interfaces. The brain learns a model of the plant to be controlled (system dynamics identification) by comparing neural commands and output (z) of the interface. New synergies are developed through controller design based on the system identified, which are then utilized while adjusting neural commands.