Scientists have developed a robotic butterfly that uses artificial intelligence to discover wing shapes that dramatically improve flight performance. The optimized designs increased lift by 339% and thrust by 46%, potentially paving the way for more efficient insect-inspired flying robots.
Butterflies are among nature’s most graceful flyers, capable of moving through the air with remarkable agility. Their flight depends on several factors, including wing shape, wing movement, body size and the surrounding environment. Although scientists have long studied how butterfly wings influence flight, identifying the precise relationship between wing geometry and aerodynamic performance has remained challenging.
Now, a research team led by Haifeng Huang has developed a robotic butterfly that allows scientists to investigate these relationships under controlled experimental conditions. By combining robotics, motion-capture technology and deep reinforcement learning, the researchers identified wing designs that significantly improved the robot’s ability to generate lift and thrust.
The findings could help scientists better understand the principles behind butterfly flight while providing new ideas for designing small, agile flying robots.
Why Butterfly Wing Shape Matters
Butterflies have evolved different wing shapes to survive in diverse environments. Their wings vary in size, proportions and overall outline, influencing how they move through the air.
These differences can affect how much aerodynamic force a butterfly generates, how efficiently it flies and how it responds to changes in its flight direction.
However, studying these relationships in living butterflies presents several difficulties. Real butterflies cannot be instructed to repeat exactly the same flight path during every experiment. Their movements may change depending on environmental conditions, lighting, individual behavior and other factors.
Consequently, researchers may struggle to determine whether a difference in flight performance results from wing shape or from changes in the butterfly’s behavior.
Previous studies have also used static butterfly-wing models in wind tunnels. Although these models help scientists investigate aerodynamic forces, they cannot fully reproduce the complex wing movements and body oscillations associated with actual butterfly flight.
Computer simulations provide another method for studying wing designs, but their predictions can differ from the performance of physical flying machines.
To overcome these limitations, the research team developed a controllable robotic butterfly that combines realistic wing movements with measurable flight performance.
Meet USTButterfly-II: A Robotic Butterfly
The researchers designed a robotic butterfly called USTButterfly-II to serve as an experimental platform for studying wing morphology.
The robot uses a motor connected to a mechanical system consisting of double-crank and double-rocker linkages. This mechanism drives both wings in synchronized flapping movements, reaching a frequency of up to five cycles per second, comparable to the flapping frequency of some real butterflies.
The robot can also be controlled wirelessly through Bluetooth, allowing researchers to operate it during flight experiments.
Its artificial wings are constructed using carbon-fiber and glass-fiber rods that imitate the veins found in natural butterfly wings. These lightweight structures also allow the wings to deform passively during movement, helping reproduce some characteristics of biological wings.
Most importantly, the researchers can modify the geometry of the artificial wings while keeping the experimental platform relatively consistent.
This gives them a major advantage: they can investigate how changing wing shape affects flight without relying on the unpredictable behavior of living insects.
How Scientists Tested Different Wing Shapes
To understand the relationship between wing geometry and flight performance, the researchers developed a collection of artificial butterfly wings with different shapes.
Rather than relying on a single measurement, they used three butterfly wing morphological parameters to describe and modify the wing geometry. These parameters allowed the team to explore different designs systematically.
The researchers also analyzed the shapes of 172 butterflies belonging to the family Papilionidae, which includes swallowtail butterflies.
Their analysis showed that the artificial wings had a high degree of morphological similarity to wings found in the Parides group of butterflies.
This comparison helped establish a biological reference for the robotic designs.
The team then used a motion-capture system to track the robot’s movements during flight. By analyzing its flight trajectories, they examined whether the robot reproduced important characteristics of butterfly-like flight.
They also defined five flight characteristic parameters to measure performance quantitatively. These measurements helped the researchers compare how different wing designs influenced the robot’s movement and aerodynamic capabilities.
Statistical analysis, including analysis of variance, showed that all three wing morphology parameters had significant effects on flight performance.
The results demonstrated that wing geometry was an important factor in determining how effectively the robotic butterfly flew.
Artificial Intelligence Searches for Better Wings
After collecting experimental data, the researchers turned to artificial intelligence to identify improved wing designs.
They employed deep reinforcement learning, a machine-learning approach in which an artificial agent learns to make better decisions by evaluating the outcomes of its actions.
In this experiment, the objective was to discover wing geometries that maximized two different aspects of flight performance: lift and thrust.
Lift is the aerodynamic force that helps an aircraft remain airborne, while thrust is the forward-directed force that helps propel it through the air.
These forces serve different purposes. A flying robot designed to carry a small payload may require greater lift, whereas a robot intended to move quickly may benefit from greater thrust.
The team first developed a neural network that acted as a substitute for the experimental environment. This model learned the nonlinear relationship between wing geometry and the measured flight characteristics, allowing the AI system to evaluate potential designs without physically testing every possibility.
The researchers also introduced additional wing variations to expand the dataset and cover a wider range of possible shapes.
They then used a deep Q-network, or DQN, algorithm to train the AI agent to search for better designs.
A particularly important feature was a multiple-feedback mechanism that helped guide the learning process. It incorporated information about optimization objectives across multiple steps and included a boundary-buffer mechanism intended to encourage further exploration of the design space.
According to the study, this approach improved the agent’s optimization performance by 33% for lift and 11% for thrust compared with a traditional reinforcement-learning feedback strategy.
This made the AI system more effective at identifying promising wing geometries.
Lift Increases by 339% and Thrust by 46%
The most striking results came from comparing the optimized wings with the baseline design.
The AI-generated wing optimized for lift achieved a 339% increase in lift performance, while the design optimized for thrust delivered a 46% increase in thrust performance.
These improvements demonstrate how substantially changing wing geometry can affect the performance of a flapping-wing robot.
The researchers also tested the optimized designs through physical flight experiments. The relative errors between the model’s predictions and the experimental results remained below 7%.
This close agreement suggests that the learned model was effective at predicting the performance of the tested designs within the experimental setup.
However, the improvements should be interpreted in context. The reported percentages describe performance relative to the study’s baseline robotic wing, rather than a 339% increase in overall flight speed or a 339% improvement in every aspect of flight.
The results also do not establish that the same numerical improvements would occur in living butterflies.
Nevertheless, the experiments demonstrate the potential of combining physical robotics with machine learning to optimize complex biological designs.
Potential Applications of Butterfly-Inspired Robots
Small flying robots inspired by butterflies could eventually find applications in several fields.
One possibility is covert reconnaissance, where small, maneuverable robots could help collect information in environments that are difficult for conventional drones to access.
Another potential application is search and rescue. Compact flying machines could assist in inspecting damaged buildings, navigating confined spaces or gathering information about hazardous locations.
Improved lift could also be useful when a robotic butterfly needs to carry sensors or other lightweight equipment. Greater thrust could support different flight requirements, although practical performance would also depend on factors such as power consumption, stability and control.
The broader research method could prove useful beyond butterfly-inspired machines. By combining controllable physical experiments with AI-based optimization, engineers may be able to investigate other flapping-wing designs and identify configurations that are difficult to discover through conventional trial and error.
Important Limitations and Future Research
Despite the promising results, the robotic butterfly cannot reproduce every aspect of a living butterfly.
Real butterflies have complex wing flexibility, biological structures and interactions between their wings and bodies. Differences in scale and flight mechanics also mean that the robotic system does not perfectly replicate the aerodynamic conditions experienced by real insects.
For this reason, the researchers emphasize that the relationships identified in their experiments should be treated as testable biological hypotheses rather than definitive explanations of how all butterflies fly.
Further research will be needed to compare these findings with detailed measurements of real butterfly wings and observations of their flight movements.
Another limitation is that the study examined the independent effects of individual morphological parameters without fully investigating how those parameters interact. In nature, several aspects of wing geometry can influence flight simultaneously.
Future experiments involving multiple parameters at the same time could reveal additional relationships and help researchers develop more realistic wing designs.
The team also plans to improve the robot’s biological resemblance, including its wing flexibility, scale and flapping mechanism.
Conclusion
The development of USTButterfly-II demonstrates how robotics and artificial intelligence can work together to investigate the relationship between biological structure and physical performance.
By testing different artificial wing shapes, measuring their flight characteristics and using deep reinforcement learning to optimize their geometry, researchers achieved substantial improvements in lift and thrust.
Although the findings cannot yet be directly applied to living butterflies, the approach offers a powerful way to generate new scientific hypotheses and refine bio-inspired flying machines.
The research highlights an important principle in engineering: sometimes, improving a flying robot does not require a more powerful motor or a larger battery. Finding a better wing shape may be just as important.
Reference: Huang, H., Chen, Z., Yang, LJ. et al. Bio-inspired exploration of butterfly wing morphology guides deep reinforcement learning optimization of robotic flight. npj Robot 4, 52 (2026). https://doi.org/10.1038/s44182-026-00116-w
Comments
Post a Comment