Distributed Non-Communicating Multi-Robot Collision Avoidance via Map-Based Deep Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Multi-robot systems face challenges in distributed, communication-free environments, especially with varying robot shapes.
• Existing collision avoidance methods often require inter-robot communication or detailed movement data, limiting their applicability.

Purpose of the Study:

• To develop a robust and efficient multi-robot collision avoidance system for distributed, communication-free scenarios.
• To leverage deep reinforcement learning with egocentric local grid maps for enhanced situational awareness and control.

Main Methods:

• A map-based deep reinforcement learning approach using egocentric local grid maps to represent the environment.
• Distributed Proximal Policy Optimization (DPPO) to train a convolutional neural network mapping grid maps to control commands.
• Multi-stage curriculum learning in simulation to improve performance before real-world deployment.

Main Results:

• The proposed method demonstrates robustness to noisy sensor data and does not require robots' movement data.
• Experiments in simulation and on real differential-drive robots show efficient and superior performance compared to existing DRL-based approaches.
• Ablation studies confirm the benefits of egocentric grid maps and multi-stage curriculum learning.

Conclusions:

• The map-based deep reinforcement learning approach offers an efficient and deployable solution for multi-robot collision avoidance in challenging environments.
• The method effectively handles varying robot shapes and sensor limitations without communication.
• This work advances the field of autonomous multi-robot navigation through innovative use of environmental mapping and reinforcement learning.