Supplementary work for AAAI-2016 Student Abstract
|Tyler M. Frasca||Antonio G. Sestito||Craig Versek|
|Douglas E. Dow||Barry C. Husowitz||Nate Derbinsky|
The video above shows end-to-end operation using kNN and Electromyography (EMG) signals to control a robotic arm. The performance, accuracy and test time, of several supervised learning algorithms were compared using Weka 3.6. The results showed kNN, k of 4, performed the best with an accuracy of >97% and a test speed of <1ms. The delay between the operator’s response and the robot’s response was purposefully implemented to reduce the jittery motion of the robotic arm.
kNN 10-fold Cross Validation
The table above compares the accuracy of 10-fold cross validation, 369 instances, using k values from 1-10. All of the k values perform >97% but 1,2 and 4 perform the best. We chose k of 4, because a small k value e.g. 1 or 2 will be susceptible to noise.
Action potentials from the four muscle groups (Bicep, Tricep, Anterior Deltoid, Posterior Deltoid) were sampled and amplified by an amplification board developed by NeuroFieldz. The PC applies a bandpass filter and then sends the data to the kNN algorithm. A block diagram of the system is shown in Figure 1.
Below is a brief description of each component in the block diagram.
In order to test the validity of end-to-end operation, we used the PhantomX Reactor Research Robot Arm Kit (Trossen Robotics). The robotic arm kit is comprised of 7 Dynamixel AX-12a robotic actuators, which have a stall torque of 1.8 Newton meter and a resolution of 0.29 degrees. Each Dynamixel actuator is equipped with a dedicated microcontroller with numerous capabilities, two important features are positional feedback and a no torque mode. The no torque mode is important for acquiring the training output positions, because the robotic arm can be positioned to the mirror the operator’s arm and still provide positional feedback.
Acquiring low-noise data is important for EMG signal acquisition because the voltage amplitude may range from a few microvolts to a few millivolts (Basmajian and De Luca 1985). The EMG amplifier board is based on the low-noise ADS1299 chip specifically made for biopotential measurements, which was an early stage prototype developed by NeuroFieldz. Each channel of the ADS1299 is a differential amplifier measuring the action potential of a single muscle. The chip is also capable of simultaneously sampling multiple channels at 1000 samples per second.
Silver/Silver Chloride (Ag/AgCl) is commonly used for electrodes, because they are non-polarized allowing current to pass across the interface of the skin and electrode with ease. This is important when dealing with motion artifacts, which can increase the amount of charge stored between the interfaces (Lee, Stephen, Kruse, and Analog Devices). Covidien Ag/AgCl electrodes (Kendall 850 Foam Electrodes) were used as the electrical transducer for the muscular impulse, since they are used in the medical field for conducting electrocardiograph tests.
In order to capture enough information for the machine learning algorithm, preliminary testing comprised of recording four muscles on the right arm with EMG. The four EMG signals became the attributes for the machine learning algorithm. Complementary muscles were chosen in order to have one muscle as the agonist muscle while the others act as the antagonist muscles. Two muscles pairs which complement one another respectively are the bicep and tricep as well as the anterior and posterior deltoids. The electrodes were placed in line with the muscle Fiber (Florimord) to allow for the best data acquisition. The pair of electrodes on one muscle were separated about 3 cm apart from one another. If the electrodes are placed over the middle of the muscle, every time the muscle is flexed the electrodes will move and the signal is susceptible to motion artifacts. Motion artifacts are one of the major issues with EMG recording and requires filtering.
Figures 2-4 show the three muscle positions used for training/testing points. The three positions allowed us to use a supervised learning algorithm, because we were able to map the EMG signals to specific position.
Figures 5-8 show the mean voltage potentials for each muscle. For each figure, the colors represent the three output positions.
Wentworth Institute of Technology & Accelerate
Northeastern University -Sridhar Lab
Basmajian, JV and De Luca, CJ. Muscles Alive (5th edition), Williams and Wilkins, Baltimore, MD, 1985.
Lee, Stephen, John Kruse, and Analog Devices. “Biopotential Electrode Sensors in ECG/EEG/EMG Systems.” Print.
Florimond, V. “Basics of SURFACE ELECTROMYOGRAPHY Applied to Physical Rehabilitation and Biomechanics.” Print.