Estimating 6-DOF Force/Torque based on the Force Myography - PowerPoint Presentation

Slide1

Estimating 6-DOF Force/Torque based on the Force Myography (FMG) Signals: A Comparative Study

Simon Fraser University

Maram Sakr,

Alaa

Eldin

Abdelaal, Syed Tanveer Jishan, and Omar Eltobgy

Simon Fraser University

MOTIVATION

METHODOLOGY

PROPOSED APPROACH

RESULTS

CONCLUSIONS

Applied research has shown that efficiency, flexibility, and quality in automated manufacturing plants, such as automotive assembly lines, can be highly improved through close cooperation between workers and robotic manipulators.

Workers operating and maintaining automated machinery are at risk of serious injuries. US statistics suggest that 18,000 amputations and over 800 fatalities in the United States each year are attributable to such causes.

Furthermore, 4.9% of wrist and hand injuries involve amputations with more than one in ten involving a manufacturing employee.

The most common cause of work related injury is exposure to inanimate mechanical forces, which accounts for 46% of work related hospitalizations with the most common bodily location being the wrist or hand (38%).

We consider the handover task between the robot and human.

We estimate the hand force and torque in 6 degree-of-freedom (DOF) using Force Myography (FMG) signals.

The FMG signals represent the volumetric changes in the arm muscles due to muscle contraction or expansion.

A force value approaching threshold limits is an indication of an approaching hazardous situation.

OBJECTIVES

We want to answer three main questions

in this study

:

Is it

feasible

to use FMG signals to predict the exerted force/torque by the human hand in

6 Degrees of Freedom (DOF)

?

What is the

best placement of the FMG sensors on the human arm?How many FMG sensors do we need?

Four force-sensing bands were placed on the participant arm on the wrist, forearm midway, forearm muscle belly and the upper arm. They were used to collect the FMG signals during force/torque exertion.An accurate force/torque sensor was used to label the FMG data.The data was collected during 5 sessions. In each session, the participant exerted force/torque in 6 DOF.The resultant force/torque values are forming an approximate sinusoidal wave in each direction.

PROPOSED EXPERIMENTS

To achieve our three objectives, we perform the following experiments:

For studying the feasibility:

we compare between different regression algorithms.

For exploring the best number and placement of the sensors: we repeat the above experiment for all possible combinations of the four bands.

WHY MACHINE LEARNING?

No first principle laws to map between the FMG signals and the exerted force.

The only viable option is to use data-driven approach.

It is a multi-output regression problem. 38 inputs are used to predict 6 outputs representing the 6 DOF force/torque.One model is built to predict the 6 DOF force/torque all together to capture the correlations between the outputs.

We tried 4 different machine learning methods on different combination of bands.For one band, MLP outperformed the other algorithms and the position of bands 1 and 3 provided the best results. Using two bands combination, KRR and GRNN performed almost similarly, while MLP outclassed those in two particular instances. The combination of bands 1 and 3 gave the best performance across all combinations.When we used three bands, GRNN provided better accuracy for every different combinations and this behavior remained consistent when all four bands were used.

The presented study explored the viability of using the FMG on the arm for 6-axis hand force/torque estimation, and also examined the effects of the FMG sensors density and location to the force/torque estimation.We performed a comparative study between four regression algorithms including kernel ridge regression (KRR), general regression neural network (GRNN), Multilayer Perceptron (MLP) and MLP with AutoEncoders.The results showed that FMG achieves a good performance in multiple-DOF force/torque estimation, with an average R2 accuracies of 0.82 across the 6 outputs using four FMG bands on the arm in cross-session evaluation. The findings from this study confirm the viability of using the FMG signals from the arm for multiple axes isometric hand force and torque around the wrist estimation, and knowledge gained from this study provides guidance for hand force/torque estimation in terms of optimal FMG sensors location and density.

Force/Torque Sensor

FMG Sensors

Battery

Arduino

R

2

Accuracy