Non-Communication Decentralized Multi-Robot Collision Avoidance in Grid Map Workspace with Double Deep Q-Network

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Multi-robot systems require efficient collision avoidance strategies.
• Decentralized control is crucial for scalability in large-scale robot deployments.
• Processing sensor data directly, like Lidar signals, can reduce communication overhead.

Purpose of the Study:

• To develop a novel decentralized multi-robot collision avoidance method using deep reinforcement learning.
• To enable robots to navigate large-scale grid map workspaces efficiently.
• To minimize directional changes while ensuring collision-free movement.

Main Methods:

• Utilized deep reinforcement learning, specifically the Double Deep Q-Network (DDQN) algorithm.
• Developed a handcrafted reward function prioritizing collision avoidance and minimal direction changes.
• Trained the policy in a simulated grid map workspace environment.
• Directly processed Lidar signals for robot control, eliminating inter-robot communication.

Main Results:

• The learned policy effectively guided robots from initial to goal positions in the grid map.
• The method successfully achieved collision avoidance among multiple robots during navigation.
• Demonstrated the system's capability in large-scale grid map environments.

Conclusions:

• The proposed decentralized deep reinforcement learning method is effective for multi-robot collision avoidance.
• Direct Lidar signal processing offers a viable alternative to communication-based methods.
• The approach is suitable for large-scale grid map workspaces, balancing navigation efficiency and safety.