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

Abstract

The challenge of flocking collision avoidance lies in selecting the optimal strategy that balances decision-making intelligence and resource saving in a complex dynamic environment. Meanwhile, the design of the collision avoidance reward function for deep reinforcement learning (DRL) in this scenario may lack a quantitative basis. To overcome this challenge, this paper proposes an event-triggered proximal policy optimization (ETPPO) avoidance strategy by combining rewards from a dynamic model under ideal environments and DRL rewards. Firstly, this strategy incorporates the intermittent communication costs of event-triggered mechanisms (ETM) to achieve a balance between avoidance performance and communication resources. Secondly, a composite avoidance reward mechanism is designed, which combines the cost function based on dynamic model and obstacle avoidance reward based on DRL. The quality and rationality of reward function design in complex environments are improved. Then, to make full use of historical information and focus on task-related key status information, an LSTM-Attention (LA) fusion module combining long short-term memory (LSTM) and attention mechanisms is introduced, and the ETPPO-LA algorithm is constructed. The stability of the network and training efficiency of the algorithm are improved. Finally, the proposed algorithm is verified based on the Ros-Stage simulation platform, which shows the advantages in terms of accumulated rewards and avoidance success rate.