Self-organizing maps and UAV Swarm morphing 

Self-Organizing Maps (SOMs) are a type of unsupervised neural network that project high-dimensional data onto a lower-dimensional (typically 2D) grid while preserving topological relationships. In the context of UAV swarm transformations and decentralized coordination, SOMs offer a powerful framework for enabling autonomous behavior, spatial awareness, and adaptive decision-making across distributed agents. The trajectory of a UAV swarm much like that of a standalone drone is specified by waypoint selection and arrival time calculation with specific speed up and speed down between waypoints to fit the schedule. The choice of waypoints is usually decided by controllers or cloud computing by determining nodes representing 3D grids in space between start and finish and edges representing reachability. Usually, it is assumed that there is enough clearance for a UAV swarm to follow the trajectory but dynamic wind turbulence, minimum distance between units and obstructions such as when passing under bridges or through tunnels require UAV swarm transformations. Just like waypoint selection, the obstructions can either be negotiated or avoided in a centralized or decentralized manner usually with some form of error correction to revert to the planned trajectory. The Hungarian algorithm computes the minimum distance transformations for the drones in a centralized manner, but SOM helps to do so in a decentralized manner. The results of SOM can be used to augment the assignments for the units in the drone or for cross-validation of alternative dynamic transformations, if not completely, eliminate the need. 

SOMs consist of a grid of neurons; each associated with a weight vector of the same dimension as input data. When an input vector is presented, the neurons whose weight vector is closest as determined by Euclidean distance, becomes the Best Matching Unit. The BMU and its neighbors update their weights to become more like the input, reinforcing spatial clustering. As with any neural network, there is training involved, and the input data is iteratively fed. In each iteration, the BMU is found, the weights of the BMU and its neighbors are updated, and the learning rate and neighborhood radius are reduced over time. This results in a map where similar inputs are clustered together, preserving the data’s topology. 

This can be applied to a UAV swarm for dynamic collision avoidance from external threats. If the units of a UAV swarm are flying sequentially in a linear path and an external object is going to intrude into the path, the resulting boundary of the object on the flight path can be fed as input vector with sufficient margin from the boundary. Then the units will find their positions on the modified path and continue to follow each other sequentially until the swarm clears the obstruction. The roles and spatial positions of the UAV swarm units can be mapped; the formation can be adapted based on terrain or task density and the node failures can be handled gracefully by reassigning roles. 

Use of SOM can go beyond decentralized coordination to include role assignment and task clustering such as to learn optimal task distributions, mapping swarm participants to mission zones, and dynamically reconfiguring based on environmental feedback. It can also be used to improve scalability and fault tolerance.  

Sample application: https://github.com/ravibeta/Node-Element-Predictions  

#codingexercise: CodingExercise-08-10-2025.docx