Bio-inspired exploration of butterfly wing morphology guides deep reinforcement learning optimization of robotic flight

How wing morphology shapes flight performance is a central question in functional biology and bio-inspired robotics, yet these relationships are difficult to isolate in living organisms. Here, we use a robotic butterfly with tunable wing geometry to systematically characterize morphology–performance relationships by combining motion capture with morphometric analysis. We then use deep reinforcement learning to search the morphological design space and identify wing designs optimized for aerodynamic performance. The resulting designs increase lift and thrust performance by 339% and 46%, respectively. Physical flight tests validate the learning-derived morphology, with prediction errors remaining below 7%. By integrating controlled robotic experiments with machine learning, this framework provides a means to generate testable hypotheses in functional morphology while guiding the engineering design of bio-inspired robots.