Fused mechanomyography and inertial measurement for human-robot interface

Abstract

Human-Machine Interfaces (HMI) are the technology through which we interact with the ever-increasing quantity of smart devices surrounding us. The fundamental goal of an HMI is to facilitate robot control through uniting a human operator as the supervisor with a machine as the task executor. Sensors, actuators, and onboard intelligence have not reached the point where robotic manipulators may function with complete autonomy and therefore some form of HMI is still necessary in unstructured environments. These may include environments where direct human action is undesirable or infeasible, and situations where a robot must assist and/or interface with people. Contemporary literature has introduced concepts such as body-worn mechanical devices, instrumented gloves, inertial or electromagnetic motion tracking sensors on the arms, head, or legs, electroencephalographic (EEG) brain activity sensors, electromyographic (EMG) muscular activity sensors and camera-based (vision) interfaces to recognize hand gestures and/or track arm motions for assessment of operator intent and generation of robotic control signals. While these developments offer a wealth of future potential their utility has been largely restricted to laboratory demonstrations in controlled environments due to issues such as lack of portability and robustness and an inability to extract operator intent for both arm and hand motion.

Wearable physiological sensors hold particular promise for capture of human intent/command. EMG-based gesture recognition systems in particular have received significant attention in recent literature. As wearable pervasive devices, they offer benefits over camera or physical input systems in that they neither inhibit the user physically nor constrain the user to a location where the sensors are deployed. Despite these benefits, EMG alone has yet to demonstrate the capacity to recognize both gross movement (e.g. arm motion) and finer grasping (e.g. hand movement). As such, many researchers have proposed fusing muscle activity (EMG) and motion tracking e.g. (inertial measurement) to combine arm motion and grasp intent as HMI input for manipulator control. However, such work has arguably reached a plateau since EMG suffers from interference from environmental factors which cause signal degradation over time, demands an electrical connection with the skin, and has not demonstrated the capacity to function out of controlled environments for long periods of time.

This thesis proposes a new form of gesture-based interface utilising a novel combination of inertial measurement units (IMUs) and mechanomyography sensors (MMGs). The modular system permits numerous configurations of IMU to derive body kinematics in real-time and uses this to convert arm movements into control signals. Additionally, bands containing six mechanomyography sensors were used to observe muscular contractions in the forearm which are generated using specific hand motions. This combination of continuous and discrete control signals allows a large variety of smart devices to be controlled.

Several methods of pattern recognition were implemented to provide accurate decoding of the mechanomyographic information, including Linear Discriminant Analysis and Support Vector Machines. Based on these techniques, accuracies of 94.5% and 94.6% respectively were achieved for 12 gesture classification. In real-time tests, accuracies of 95.6% were achieved in 5 gesture classification.

It has previously been noted that MMG sensors are susceptible to motion induced interference. The thesis also established that arm pose also changes the measured signal. This thesis introduces a new method of fusing of IMU and MMG to provide a classification that is robust to both of these sources of interference.

Additionally, an improvement in orientation estimation, and a new orientation estimation algorithm are proposed. These improvements to the robustness of the system provide the first solution that is able to reliably track both motion and muscle activity for extended periods of time for HMI outside a clinical environment.

Application in robot teleoperation in both real-world and virtual environments were explored. With multiple degrees of freedom, robot teleoperation provides an ideal test platform for HMI devices, since it requires a combination of continuous and discrete control signals. The field of prosthetics also represents a unique challenge for HMI applications. In an ideal situation, the sensor suite should be capable of detecting the muscular activity in the residual limb which is naturally indicative of intent to perform a specific hand pose and trigger this post in the prosthetic device. Dynamic environmental conditions within a socket such as skin impedance have delayed the translation of gesture control systems into prosthetic devices, however mechanomyography sensors are unaffected by such issues.

There is huge potential for a system like this to be utilised as a controller as ubiquitous computing systems become more prevalent, and as the desire for a simple, universal interface increases. Such systems have the potential to impact significantly on the quality of life of prosthetic users and others.