Deep Reinforcement Learning for Intelligent Frontier Ranking in Search-and-rescue Scenarios

Bergeron, Taylor

Etd

Deep Reinforcement Learning for Intelligent Frontier Ranking in Search-and-rescue Scenarios

Public

Robot-based search-and-rescue is a compelling application due to its dangerous and time-critical nature. However, exploring real-world environments efficiently is difficult because of their complex and dynamic features, and of the need to pick the most promising unexplored areas. This research focuses on reinforcement learning for ranking frontiers (i.e., unexplored areas) for a search-and-rescue application in an indoor environment. My approach is based on Advantage Actor Critic (A2C), a reinforcement learning method which combines two prominent reinforcement learning (RL) algorithms. I implement the method using Robot Operating System (ROS) and the Gazebo simulator, and the OpenAI Gym to setup the reinforcement learning environment. I train a singular agent to find the optimal point to navigate to in a given environment, with the goal of exploring the environment as quickly as possible. Then, I implement the model on each robot in a swarm, to have the swarm explore the environment as quickly as possible. Experimental evaluation shows that my approach reduces exploration time by 8.4% with respect to traditional approaches. In addition, the software I developed in this project is the first to integrate popular SLAM ROS packages and the Gazebo simulator with OpenAI ROS Gym, providing the means to further this line of research with more sophisticated approaches.

Creator