Intention recognition of UAV swarm with data-driven methods
image: Unmanned Aerial Vehicle (UAV) intention recognition has been extensively studied. However, when it comes to a UAV swarm problem, their intention recognition study was left blank. The flowchart demonstrates the framework of the intention recognition method proposed in this research. To describe the UAV swarm's basic motions, the behaviors of a UAV swarm are classified into contraction, free movement, and expansion. Once the data are collected from UAVs and stored in an array, an Artificial Neural Network (ANN) can detect the motion and analyze the probabilities of contraction, free movement, or expansion motion.
Credit: Zhichao Wang, Jiayun Chen, Jiaju Wang, Qiang Shen.
Single Unmanned Aerial Vehicle (UAV) intention recognition has been well studied. When there are more than three UAVs, traditional recognition may not precisely describe the UAVs’ behavior. Researchers from Shanghai Jiao Tong University have developed a novel data-driven approach to recognize the intentions of unmanned aerial vehicle (UAV) swarms, addressing a critical gap in aerial surveillance and defense systems. While single UAV intention recognition has been extensively studied, the ability to interpret coordinated behaviors among multiple drones has remained a significant challenge.
Traditional recognition methods fail to capture the nuanced collective behaviors when dealing with three or more UAVs operating as a coordinated swarm. New approach treats the swarm as a whole intelligent entity rather than individual units, which is crucial for accurate intention prediction.
Rather than relying solely on theoretical models, the researchers combined the Dubins motion model with an artificial neural network to create a practical intention recognition system. The Dubins model, a simplified representation of UAV movement, provided a realistic foundation for generating training data while maintaining computational efficiency.
The special experimental scenario this article is interested in is a three-agent triangle motion problem. The problem is to recognize the three basic intentions before the movement is half done by the drones. The research team identified that among 63 different motion parameters they initially considered, just two measurements, linear acceleration and rotation angle relative to the swarm's centroid, provided the most reliable prediction.
Back to the flowchart, the system employs a multi-step process to analyze swarm intentions. It first checks if the swarm is in motion, then collects velocity, acceleration, and position data from each UAV. The information is accumulated and averaged over time to filter out noise and environmental deviations before being processed by a specifically designed Multi-Layer Perceptron neural network, which identifies the swarm's intention.
The researchers validated their approach through both simulation and real-flight testing. Using Microsoft's AirSim platform built on Unreal Engine 4, they generated realistic flight data under controlled conditions. More importantly, they conducted actual flight experiments using DJI Tello UAVs, manually guiding the drones along precisely marked paths while the system continuously analyzed their movement.
The results were impressive: the system achieved over 98% accuracy in intention recognition throughout the entire movement process of the UAV swarm, with unstable outputs accounting for less than 15% of the time. Crucially, the method could predict swarm intentions within the first half of each movement sequence, giving decision-makers critical early warning time.