Depth-based obstacle avoidance through deep reinforcement learning

• Main Authors: Wu, Keyu, Mahdi Abolfazli Esfahani, Yuan, Shenghai, Wang, Han
• Other Authors: School of Electrical and Electronic Engineering
• Format: Conference or Workshop Item
• Language: English
• Published: 2020
• Subjects: Engineering::Electrical and electronic engineering; Deep Reinforcement Learning; Obstacle Avoidance
• Online Access: https://hdl.handle.net/10356/142309
• Institution: Nanyang Technological University

Obstacle avoidance is an indispensable technique for mobile robots to maneuver safely without collision. In this paper, we propose an end-to-end deep neural network to derive control commands directly from the raw depth images using deep reinforcement learning. The convolutional neural networks are used to extract the feature representation from the input depth images and the fully connected neural networks subsequently map the features to Q-values for determination of the optimal action. To improve the performance of the network, we adopt a two-stage method so that noisy fully connected layers are employed at the beginning while conventional ones are utilized during the second stage of training. Compared to the existing method, our proposed model exhibits much better performance in avoiding obstacles and converges faster during training.