Robotic Navigation with Human Brain Signals and Deep Reinforcement Learning

摘要

Navigation under a grid world has been a classical and historical theme in reinforcement learning, which is viewed as a Markov decision process (MDP) in general. The literature to date has proven that the navigation task can be addressed that an agent achieves the target with a high success rate and avoids collision with obstacles. But essentially, they met with success in a specific environment while the agent cannot work in a new surrounding, meaning that it does not boast broad applicability. In state-of-the-art approaches, poor feedback and lack of adaptability to increasing state spaces remain a problem. In this paper, we propose a modified approach to solve a series of navigation problems under moderate and huge-sized surroundings. The problem is addressed with a deep reinforcement learning algorithm with a guided classifier. We address these issues by providing a reliable guided reward with a brain-guided classifier based on human brain signals (electroencephalography, EEG) and a convolutional neural network. This paper explores several experiments to show that our model with deep RL and the brain-guided classifier can solve these complex and significant practical challenges. Our method improves efficiency by about twice as much as traditional approaches such as DQN and Q-learning.