Natural systems such as migratory birds achieve remarkable energy efficiency through self-organization and dynamic formation reconfiguration. We show that fully decentralized, memoryless ground robots can reproduce these effects using only range and bearing sensing, a digital compass, and battery level monitoring. We apply an existing evolutionary framework capable of optimizing Hebbian plasticity parameters of neural networks, giving robots the ability to continuously adapt and learn. In a uniform headwind setting, robots learn to form drag-reducing patterns and exhibit emergent formation reconfiguration that reallocates the energetic load, based on battery levels and without relying on direct communication or any wind sensor. Validation experiments in simulation show that the resulting controller outperforms a traditional flocking baseline method. Our results show that the adaptive controller can lead to the emergence of formation reconfiguration in the presence of very limited local information.