Abstract

Hand gesture recognition (HGR) based on electromyography (EMG) has been a research topic of great interest in recent years. Designing an HGR to be robust enough to the variation of EMGs is a challenging problem and most of the existing studies have explored supervised learning to design HGRs methods. However, reinforcement learning, which allows an agent to learn online while taking EMG samples, has barely been investigated. In this work, we propose a HGR system composed of the following stages: pre-processing, feature extraction, classification and post-processing. For the classification stage, we use Q-learning to train an agent that learns to classify and recognize EMGs from five gestures of interest. At each step of training, the agent interacts with a defined environment, obtaining thus a reward for the action taken in the current state and observing the next state. We performed experiments using a public EMGs dataset, and the results were evaluated for user-specific HGR models by using a method that is robust to the rotations of the EMG bracelet device. The results showed that the classification accuracy reach up to 90.78% and the recognition up to 87.51% for two different test-sets for 612 users in total. The results obtained in this work show that reinforcement learning methods such as Q-learning can learn a policy from online experiences to solve both the hand gesture classification and the recognition problem based on EMGs.