Hierarchical Deep Reinforcement Learning for Multi-Robot Systems Application in Dynamic Environments

 

Reinforcement Learning (RL) is a method that has been developed for robot control in recent years. The motivation is how the robot can be controlled automatically, and the robot can learn independently. The basic principle of RL is about learning how robots behave of reward structure. The problems are complex and dynamic environment resulting in large amounts of data (state or action space) are to be learned. So that this problem cannot be solved with traditional RL. Deep Reinforcement Learning (DRL), a combination of deep learning and reinforcement learning, is used to overcome the weaknesses of the RL. DRL, neural networks as function approximators, is very useful for estimating value when the state or action space is too large. Another approach to overcome the weaknesses of the RL is Hierarchical RL (HRL). HRL divide a large problem into sub-problems in a hierarchy. The benefit of HRL is to increase exploration and transfer learning. Hierarchical Deep Reinforcement Learning (HDRL), the combination of DRL and HRL, expected to result in better performance.

In addition, Multi-Robot System (MRS) has been extensively researched and developed in recent years. The advantage of using MRS compared to a single robot is that it reduces time, increases efficiency and has better durability in completing tasks.

This research focuses on:

• Developing a new learning model for MRS using the HDRL method
  The robots can organize themselves automatically
  The robots can work together to explore effectively and efficiently
• Perform simulation and implementation on the real robot
  Dynamic environment
  Search and Rescue (SAR) application
  The robots used are wheeled robots