Distributed deep reinforcement learning based on bi-objective framework for multi-robot formation.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Multi-robot formation control is crucial for coordinated tasks.
• Improving generalization ability in multi-robot systems reduces training and computational costs.
• Existing methods may struggle with dynamic environments and generalization to new target positions.

Purpose of the Study:

• To develop a generalized multi-robot formation control method.
• To enhance the generalization ability of neural networks in multi-robot formation tasks.
• To present a distributed deep reinforcement learning approach for robust formation control.

Main Methods:

• Utilized a distributed deep reinforcement learning method based on the soft actor-critic algorithm.
• Employed a strategy of using different neural networks for distinct objectives to improve learning focus.
• Designed a formation evaluation assignment function tailored for distributed training.

Main Results:

• The improved algorithm achieved higher cumulative reward values compared to the original algorithm.
• Experimental results demonstrated superior maintenance of desired formations during movement.
• The proposed algorithm exhibited enhanced stability, as evidenced by control signal curves.
• The system showed better flexibility in formation control due to a rotation design in the reward function.

Conclusions:

• The proposed distributed deep reinforcement learning method effectively improves multi-robot formation generalization.
• The algorithm demonstrates robust performance in formation maintenance and adaptability to formation variations.
• The approach offers enhanced stability and flexibility for multi-robot systems in dynamic environments.