Design and implementation of an agent using reinforcement learning to the operation of a myoelectric prosthesis using EMG signals.

Abstract

In the world exists approximately 65 million amputees, 38.7% of which correspond to upper limb amputations, and most of them are found in developing countries. This leads to the necessity of an affordable, easy-to-made hand prosthesis to alleviate these people situation. This project proposes the use of reinforcement learning to train an agent for the operation of a myoelectric prosthesis. A prosthesis prototype with flexible materials was built from an open-source model chosen from the literature. The prothesis was modified and built as a low-cost 3D printed prototype with 4 degrees of freedom, 1 for each finger but the little. A commercial electromyography sensor is used, and just during evaluation a glove with flexion sensors. The training of the agent that operates the prosthesis was divided in two stages, simulation and fine-tuning. For this matter, a simplified model of the prosthesis dynamics was implemented for simulation. Simulation training allowed fast proof of concept testing and hyperparameters selection. Fine-tuning was necessary to adapt the agent to the real hardware. Three approaches for the reward function were considered, the best results were obtained with the combination of 2 of them: the distance to the target, and a discrete reward depending on the selected action by the agent. The performance of the prosthesis was measured using the success rate of the grasp and release tasks obtaining an 86%. From this, we conclude that this research was a successful first step towards achieving a fully functional myoelectric hand prosthesis