News Release

From black-box trial and error to glass-box flight: Physics-embedded AI for flapping-wing robots

Peer-Reviewed Publication

SciOpen

A bioinspired flapping-wing micro air vehicle developed by the Bioflight and Bioinspired Aerial Vehicles Team at Beihang University in free flight (frame from experimental video)

image: 

A bioinspired flapping-wing micro air vehicle developed by the Bioflight and Bioinspired Aerial Vehicles Team at Beihang University in free flight (frame from experimental video). In a Chinese Journal of Aeronautics Highlight perspective, Hao Li and Jianghao Wu outline how embedding unsteady fluid mechanics and energy-based stability constraints into neural control architectures can advance low-level sensorimotor autonomy.

view more 

Credit: Bioflight and Bioinspired Aerial Vehicles Team, Beihang University (Hao Li and Jianghao Wu)

As uncrewed aerial vehicles move from open skies into cluttered indoor rooms, dense forests and urban canyons, conventional quadrotors face practical physical constraints: high-speed rotor blades generate intrusive noise, pose collision hazards near people and lose aerodynamic efficiency at low Reynolds numbers (100 to 10,000). Flapping wings offer a bioinspired alternative for low-Reynolds-number flight, combining structural compliance and low-frequency acoustic signatures with unsteady vortex lift.

Over the past decade, advances in flexible wings, precision micro-transmissions and wing-abdomen coupled actuation have unlocked remarkable mechanical agility. Building on 25 years of bioflight research, the Bioflight and Bioinspired Aerial Vehicles Team at Beihang University has developed untethered hover-capable hummingbird-scale vehicles, clap-and-fling multi-wing flyers and coupled flapping-rotary platforms—including the BH-Fly prototype capable of executing a 360-degree roll in 0.18 seconds. Yet translating this airframe agility into reliable autonomous flight in real-world airflow remains a major challenge. In a Highlight perspective published in Chinese Journal of Aeronautics, Hao Li and Jianghao Wu trace this bottleneck to a mismatch between agile physical bodies and low-level flight-control intelligence.

While much of today’s embodied AI research focuses on high-level visual navigation and semantic task planning, flapping-wing micro air vehicles (FWMAVs) face a more fundamental hurdle at the level of physical and sensorimotor embodiment—how control policies couple directly with compliant wings, underactuated multibody dynamics and rapid unsteady airflow. “If the first decade of FWMAV research was defined by replicating the physical body of biological flyers, the next era will inevitably be defined by replicating their sensorimotor intelligence,” the authors write.

Both conventional model-based control and unconstrained model-free reinforcement learning face limitations in this regime. Controllers built on quasi-steady aerodynamic averages omit wake history and treat gusts merely as external disturbances to be suppressed. Conversely, pure data-driven neural networks treat the coupled fluid-structure system as an opaque “black box.” Trained through trial and error without aerodynamic or energetic priors, black-box policies require extensive sampling and risk losing stability during aggressive maneuvers or simulation-to-real (Sim-to-Real) transfer.

To address these challenges, Li and Wu propose a shift toward “glass-box” physics-embedded intelligence across two complementary fronts.

The first front models transient aerodynamics so controllers can anticipate and harness strong unsteady aerodynamic effects rather than fighting them. At low Reynolds numbers and high angles of attack, every wingbeat generates leading-edge vortices and interacts nonlinearly with the wake shed by previous strokes. Quasi-steady models miss this spatio-temporal memory, whereas full computational fluid dynamics is too computationally expensive for iterative policy optimization. Combining lattice-Boltzmann fluid-structure interaction (LBM-FSI) simulations with differentiable sequence neural surrogates provides physically grounded gradient feedback on vortex shedding, wing elastic deformation and wake capture.

The second front embeds analytical stability constraints and structural compliance into the control architecture. Biological flyers rely on elastic wing hinges as passive mechanical filters that store and release elastic energy near resonance to buffer gust impacts, paired with neuromotor loops that dissipate disturbance energy. To translate this principle into neural network design, the authors introduce Port-Hamiltonian dynamics and passivity-based energy shaping via input convex neural networks (ICNNs). Constructing positive-definite energy potential functions with dynamic damping injection ensures intrinsic energy dissipation and analytical stability constraints. Paired with differentiable central pattern generators (CPGs) that bridge low-frequency attitude commands and high-frequency wingbeats, the controller works in harmony with structural compliance to enhance flight robustness during near-ground hovering and navigation through confined micro-environments.

The long-term goal, in the authors’ words, is to enable vehicles that “foresee and orchestrate transient aerodynamics, riding the gusts rather than merely surviving the flow.”

By bridging intelligent algorithms and physical hardware through differentiable aerodynamic surrogates and physics-embedded control architectures, this paradigm lays the foundation for “biological-level agility”—enabling robust flight through turbulent gusts and precise hovering in cluttered spaces for applications such as confined-space search and rescue and emergency inspection. Beyond flapping-wing robots, this physics-embedded blueprint scales to morphing-wing aircraft, cross-medium unmanned vehicles and self-organizing drone swarms.

 

Original Source

Hao LI, Jianghao WU*. Embodied AI for bioinspired aerial vehicles: The paradigm shift toward physics-embedded intelligence[J]. Chinese Journal of Aeronautics, 2026: https://doi.org/10.1016/j.cja.2026.104452

 

About Chinese Journal of Aeronautics 

Chinese Journal of Aeronautics (CJA) is an open access, peer-reviewed international journal covering all aspects of aerospace engineering, monthly published by Elsevier. The Journal reports the scientific and technological achievements and frontiers in aeronautic engineering and astronautic engineering, in both theory and practice. CJA is indexed in SCI (IF = 6.9, Q1), EI, IAA, AJ, CSA, Scopus.


Disclaimer: AAAS and EurekAlert! are not responsible for the accuracy of news releases posted to EurekAlert! by contributing institutions or for the use of any information through the EurekAlert system.