Autonomous collision-avoiding for multi-UAVs in complex dynamic environments: An event-triggered PPO approach with...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Flocking collision avoidance requires balancing intelligent decision-making with resource efficiency in dynamic environments.
• Deep reinforcement learning (DRL) reward functions for collision avoidance often lack a quantitative foundation.

Purpose of the Study:

• To propose an event-triggered proximal policy optimization (ETPPO) strategy for flocking collision avoidance.
• To improve the design of collision avoidance reward functions in complex dynamic environments.
• To enhance algorithm stability and training efficiency through an LSTM-Attention fusion module.

Main Methods:

• Incorporated intermittent communication costs using event-triggered mechanisms (ETM) for resource-performance balance.
• Designed a composite avoidance reward mechanism combining dynamic model costs and DRL rewards.
• Introduced an LSTM-Attention (LA) fusion module to process historical and key status information, creating the ETPPO-LA algorithm.

Main Results:

• The ETPPO-LA algorithm demonstrated improved network stability and training efficiency.
• Verification on the Ros-Stage simulation platform showed significant advantages in accumulated rewards.
• The proposed strategy achieved a higher avoidance success rate compared to existing methods.

Conclusions:

• The developed ETPPO-LA strategy effectively addresses flocking collision avoidance challenges.
• The composite reward mechanism and LA fusion module enhance decision-making and resource management.
• The approach offers a robust and efficient solution for autonomous systems in complex environments.