End-to-end decentralized formation control using a graph neural network-based learning method.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Traditional multi-robot cooperative control relies on sequential perception-control pipelines, leading to latency and errors.
• End-to-end learning offers a solution but faces scalability challenges in multi-robot systems.

Purpose of the Study:

• To develop a novel, scalable, decentralized cooperative control method for multi-robot formations using deep neural networks.
• To address the limitations of existing methods by integrating perception and control.

Main Methods:

• A decentralized cooperative control method using deep neural networks.
• Inter-robot communication modeled via a graph neural network (GNN).
• Learning control policies from expert demonstrations using LiDAR sensor data.

Main Results:

• The proposed method achieves scalable control policies, effective even with a fixed training number of robots.
• Demonstrated successful triangular formation behavior in multi-robot teams of varying sizes.
• Overcame processing latencies and compounding errors inherent in sequential pipelines.

Conclusions:

• The novel GNN-based end-to-end learning approach enables scalable and robust decentralized cooperative control for multi-robot formations.
• This method offers a promising alternative to traditional model-based control for complex robotic tasks.