Caltech ICRA 2026

A Narwhal-Inspired Sensing-to-Control Framework for Small Fixed-Wing Aircraft

Fengze Xie*, Xiaozhou Fan*, Jacob Schuster, Yisong Yue, Morteza Gharib

* Equal contribution

Caltech

Narwhal-inspired UAV: probes sense clean upstream flow before it reaches the airframe

Abstract

We present an end-to-end sensing-to-control pipeline for fixed-wing UAVs combining bio-inspired hardware, physics-informed dynamics learning, and convex control allocation. Inspired by the narwhal's tusk, we mount multi-hole probes upstream of the airframe to sense "clean" airflow, reducing force-estimation error by 25–30%. A control-affine model with soft left–right symmetry regularization maps probe readings and sparse wing pressures to aerodynamic wrenches. Wind-tunnel experiments show the proposed model degrades only ~12% under distribution shift versus 44% for an unstructured baseline, with 27–34% lower tracking error.

Hardware

Instrumented fixed-wing UAV and wind-tunnel setup
(A) 1.2 m wingspan aircraft with dual multi-hole probes, 7 wing pressure taps, and onboard Jetson Orin Nano. (B) Wind-tunnel rig with two-axis rotation bed, force transducer, and gust generator. (C) Multi-hole probe detail.

Method

Sensing-to-control pipeline

Pitot Tube Calibration

Data-driven mapping from normalized probe pressures to airspeed and flow angles (Va, α, β), robust across operating conditions.

Physics-Informed Learning

Control-affine model y = Aφ(o) + Bφ(o)u with soft symmetry regularizer to resolve flaperon–wing-pressure confounding.

Convex Control Allocation

Closed-form regularized least-squares optimizer balancing force tracking, smoothness, and trim adherence.

Results

Symmetry Regularization

Without regularization the model assigns ~3.8x more lift sensitivity to the left flaperon than the right. The symmetry prior restores physically consistent control effectiveness.

Control effectiveness heatmaps: with vs without symmetry regularization
Control effectiveness BT: with symmetry regularization (left) vs. without (right).

Out-of-Distribution Robustness

Unstructured RMSE grows 44% from 10 to 14 m/s. Affine Sym grows only 12%.

Wind Speed Affine Sym Affine Unstructured
7 m/s0.5210.5310.521
9 m/s0.4950.4930.484
10 m/s (train)0.5220.4500.442
14 m/s0.5920.5980.637

Estimation RMSE (N / N·m). Bold = best per row.

Force Tracking at 14 m/s

Affine Sym reduces Fz tracking RMSE by 27% vs. Affine and 34% vs. Unstructured.

Fz estimation and tracking comparison
Left: open-loop estimation. Right: closed-loop tracking. Top to bottom: Affine Sym, Affine, Unstructured.

Citation

@inproceedings{xie2026narwhal,
  title={A Narwhal-Inspired Sensing-to-Control Framework for Small Fixed-Wing Aircraft},
  author={Xie, Fengze and Fan, Xiaozhou and Schuster, Jacob and Yue, Yisong and Gharib, Morteza},
  booktitle={IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026}
}