Skip to main content
Back

Brown University's Tiny 16-Neuron Network Keeps Drones Safe in 18 km/h Winds

A learning-enhanced control barrier function keeps a 38g nano quadrotor safe in 18km/h wind by adapting to unknown disturbances online at 100Hz.

Pro

Paper, Full Access
  • Researchers propose NODE-HO-aCBF, a control barrier function that learns model residuals online using Knowledge-based Neural ODEs.
  • The hybrid controller combines a nominal analytical model with a small neural network residual, solved as a QP at 100Hz.
  • Deployed on a 38g Crazyflie 2.1 nano quadrotor, it held a 1-meter safety distance against 18km/h wind turbulence.
  • Unlike prior adaptive CBFs, no knowledge matrix assumption or extensive tuning is needed for the residual structure.
  • Beat baseline HOCBF and HO-aCBF controllers across attractive, repulsive, and time-varying disturbance simulations.
  • Trade-off: online learning introduces a brief delay before adaptation kicks in on first exposure to a new disturbance.

Safety certificates for robots are only as good as the model they trust. A new approach from researchers at Brown University and the University of Modena and Reggio Emilia keeps that trust intact when the physics gets messy, letting a 38-gram nano quadrotor hold its safety boundary against turbulent wind that would knock a naive controller straight into an obstacle.

The paper, Online Learning-Enhanced High Order Adaptive Safety Control, tackles a well-known failure mode of Control Barrier Functions (CBFs), the mathematical safety filter that has become a staple of provably safe control. CBFs give guarantees on set invariance (safety) without compromising performance, but those guarantees ride entirely on the assumed dynamics. Add a payload, a gust of wind, or downwash from a neighboring drone, and the certificate quietly becomes invalid.

Why safety filters break in the wild

A CBF works by adding a constraint to a quadratic program: whatever control you actually want to send, project it to the closest one that keeps a scalar safety function h(x) non-negative. If the true dynamics drift away from the nominal model used to compute that constraint, the projection is solving the wrong problem. Model mismatch, external forces, payload shifts, and wind disturbances all push the real system off the assumed one.

Existing fixes each carry a catch. Robust CBFs need known uncertainty bounds and end up overly conservative. Adaptive CBFs (aCBFs) assume the residual disturbance factors into a known state-dependent matrix times a constant unknown vector, which rarely holds in practice. Learning-based residual approaches like Neural-Fly work well but demand roughly 12 minutes of offline flight data under diverse wind conditions before deployment.