Previous articles on UAV swarm flight management has focused on breaking down different stages of behavior for the swarm and coming up with strategies for each of them. For instance,

UAV swarms must have different control strategies for flight formation, swarm tracking, and social foraging. Separation of stages articulates problems to solve independent of one another, but the UAV swarm cannot always be predicted to be in one or the other at all times during its flight path because the collective behavior may not always be the most optimal at all times. This is further exacerbated by the plurality of disciplines involved such as co-ordination, aggregation, network communication, path planning, information sensing and data fusion. Also, the distributed control strategy is better suited at times from centralized control strategy, and this involves the use of consensus algorithms. Even if there is a centralized steer that corrals the herd towards a focal point, the directive to switch from individual to swarm behavior and subsequent relaxation cannot be set by the steer.

On the other hand, an approach that continuously aims to optimize the flight for each of the drones and makes it smarter to respond to obstacles with deep learning always guarantees the best possible outcome for that unit. Then the swarm behavior is about forging the units by incentivizing them to behave collectively to maximize their objectives and helps to make the behavior more natural as well as dynamic. Intelligent autonomous vehicles have already demonstrated that given a map, a route and real-world obstacles detected sufficiently by sensors, they can behave as well as humans because they apply the best of computer vision, image processing, deep learning, cognitive inferences, and prediction models. It also aids the unit to have a discrete time-integrator so that it can learn what has worked best for it. With aerial flights, the map gives way to a grid that can be walked by using sub-grids as nodes in a graph and finding the shortest connected path as route to follow. Then as each unit follows the other to go through the same route, subject to the constraints of minimum and maximum distances between pairs of units and the overall formation feedback loop, the unit can adjust shape the swarm behavior required.

In this context, the leader-follower behavior is not sacrificed but just that it becomes one more input in a feedback loop to the same strategy that works individually for each drone over time and space and the collective swarm behavior desired is an overlay over individual behavior that can be achieved even without a leader. Simulations, sum of square errors and clustering can guarantee the swarm behavior to be cohesive enough for the duration of the flight. It also enables units to become more specialized than others in certain movements so that the tasks that the swarm would have executed could not be delegated to specific units instead of all at once. Specialization of maneuvers for certain units can also come in handy to form dynamic ad hoc swarms so as to keep the control strategy as distributed, dynamic, responsive and timely as necessary. Formation of swarms and break out to individual behavior when permitted to be dynamic also results in better tracking and maximizing of objectives by narrowing down the duration over which they matter.