Researchers at the University of Science and Technology Beijing (USTB) and partner labs have shown that a deep reinforcement learning loop, fed by real flights rather than only simulation, can redesign the wings of a biomimetic flapping robot. The work appears in Nature's npj Robotics under the title describing bio-inspired exploration of butterfly wing morphology guiding deep reinforcement learning for robotic flight.
Their platform, USTButterfly-II, weighs 48.3 grams, spans 57.5 centimeters and flaps at a rated 5 hertz with an 84-degree stroke. A single motor and a double-crank, double-rocker linkage drive both wings in sync. Carbon-fiber and glass-fiber veins carry a 0.03-millimeter CPE membrane that deforms passively. After several tethered iterations, the team flew the robot untethered under wireless control inside a room tracked by 14 Optitrack cameras.
Before learning began, they encoded each wing with three morphological parameters and built 19 pairs of shapes that sit inside the natural range of 172 swallowtail specimens they measured, including several Parides species. Each pair flew five times, for 95 flights in total. Five flight metrics, including climb angle and speed, quantified performance. Sharper tips favored lift; rounder, more outboard shapes favored thrust; longer forewings helped climbing.
A neural surrogate first learned to map wing parameters to those metrics, reaching correlation coefficients of 0.97 on training data and 0.9 on held-out flights. A deep Q-network then searched the three-dimensional shape space. Multistep memory and a boundary buffer kept the agent from stalling at the edges of the design box; ablations credit those tricks with 33 percent and 11 percent better lift and thrust optimization. The lift-focused design raised the paper's lift performance parameter from 2.23 to 9.8, a 339 percent gain, and produced a 32.7-degree mean climb angle on the real robot. The thrust-focused design raised its thrust parameter by 46 percent and average speed from 1.98 to 2.92 meters per second in the reported trials. Physical checks stayed within about 7 percent of the model's predictions.
The authors stress that the numbers are performance parameters on this robot, not a claim that living butterflies gained the same lift, and that single-factor wing trials leave multi-parameter interactions for later work. Still, the closed loop from ninety-five real flights to a verified new wing offers a concrete design tool for flapping robots that need either climb and payload or speed.
SEE ALSO: Anyverse Dynamics Open-Sources MWA-Phoenix · Bionic Facial Robot Research Makes Science Robotics Cover