Decentralized Multi-Robot Navigation Based on Deep Reinforcement Learning and Trajectory Optimization.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Multi-robot systems require robust decision-making for complex tasks.
• Decentralized obstacle avoidance methods often compromise safety for computational efficiency.
• Ensuring collision-free movement is a critical challenge in multi-robot coordination.

Purpose of the Study:

• To develop an advanced algorithm for safe and efficient obstacle avoidance in multi-robot systems.
• To enhance path planning and trajectory optimization for cooperative robot navigation.
• To address limitations of existing decentralized strategies in ensuring safety.

Main Methods:

• Integration of Graph Neural Networks (GNNs) and Deep Reinforcement Learning (DRL) for path generation.
• Application of Artificial Potential Fields (APF) to incorporate safety constraints.
• Refinement of trajectories using constrained nonlinear optimization, forming the GNN-RL-APF-Lagrangian algorithm.

Main Results:

• The GNN-RL-APF-Lagrangian algorithm achieved a 96.43% success rate in sparse obstacle environments and 89.77% in dense environments.
• Demonstrated significant improvements (59-60%) over baseline GNN-RL approaches in obstacle avoidance.
• Validated scalability for systems up to 30 robots while maintaining distributed execution.

Conclusions:

• The GNN-RL-APF-Lagrangian algorithm effectively enhances safety and obstacle avoidance in multi-robot systems.
• The combined approach of APF and nonlinear optimization provides superior trajectory refinement.
• The proposed method offers a scalable and safe solution for complex multi-robot navigation challenges.