Multiagent Soft Actor-Critic Based Hybrid Motion Planner for Mobile Robots.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Multirobot systems often require explicit communication for coordinated motion planning.
• Existing planners may struggle in environments with limited observability and no communication.
• Model-free approaches offer flexibility but can be challenging to implement for complex multirobot tasks.

Purpose of the Study:

• To present a novel hybrid multirobot motion planner.
• To enable coordinated motion planning under no explicit communication and local observability.
• To achieve end-to-end mapping from multirobot states/observations to smooth, executable trajectories.

Main Methods:

• A front-end and back-end separated architecture was developed.
• The front-end utilizes the multiagent soft actor-critic (MASAC) algorithm within a centralized training with decentralized execution (CTDE) framework for collaborative waypoint searching.
• The back-end employs a minimal snap method with safety zone constraints for trajectory optimization, ensuring dynamic feasibility.

Main Results:

• The proposed planner successfully generates smooth and continuous trajectories for multiple robots.
• The system operates effectively under conditions of no explicit communication and local observability.
• Experimental results across multiple groups validated the planner's effectiveness.

Conclusions:

• The novel hybrid motion planner provides an effective solution for multirobot coordination in challenging environments.
• The model-free, end-to-end approach simplifies complex motion planning tasks.
• The combination of MASAC and minimal snap optimization ensures safe and executable trajectories.