NSM-CBAM Planner: A Brain-Inspired Autonomous Collision Avoidance Decision-Making Method for UAVs

Abstract

Autonomous collision avoidance is one of the key technologies enabling UAVs to successfully carry out measurement tasks in complex and unknown environments. Existing collision avoidance methods typically ignore the spatial and temporal differences in historical information, which makes it hard to achieve effective memory of critical historical states. To address these issues, drawing on high-level bio-inspired principles of spiking neurons and attention, a neuromorphic spiking memory planner (NSM-CBAM) is proposed in this paper. The framework is based on an actor-critic structure, consisting of the spiking memory CBAM actor network (SMCAN) and the conventional fully connected deep critic network (DCN). By introducing spiking memory CBAM into SMCAN, historical information is adaptively invoked in spatial and temporal channels, thereby achieving efficient utilization of historical memory and elimination of redundant historical information. Additionally, soft-reset leaky integrate-and-fire (SR-LIF) neurons are employed to mimic the dynamic changes in membrane potential. Spiking encoders and decoders are employed to convert between continuous states and spiking sequences. Experimental results in a Gazebo/ROS simulation environment show that, compared with other known methods, NSM-CBAM has achieved the highest collision avoidance performance in both unseen evaluation environments, with collision avoidance success rates of up to 94% and 92.5% respectively. This demonstrates the effectiveness and robustness of the proposed framework for autonomous collision avoidance in simulated outdoor environments. Simultaneously, it also provides a promising bio-inspired paradigm with potential computational efficiency for future autonomous collision avoidance of UAVs in more complex environments.