Robotics Autonomy · AirLab, CMU

Drones that feel their surroundings

Sensing nearby surfaces from aerodynamics, no extra sensors

Near a wall or floor, a drone's airflow changes and its thrust shifts. I turn that shift into distance and angle to the surface, so the drone senses without a camera or LiDAR.

< 3 cm distance error
< 4° angle error
0 sensors added

Overview

Proximity from thrust, not sensors

Research at CMU's AirLab. A drone has no cheap way to sense a surface it isn't looking at. But ground and wall effects deform its thrust and control signals in a structured way. I model that mapping to estimate distance and incidence angle, turning the propulsion system itself into a proximity sensor.

Fig 1 — As the drone nears a surface (right), the thrust signature lifts and deforms. That deviation is the signal.
01

Physics-informed

Distance and angle estimation posed as a learning problem over thrust, RPM, and IMU time-series, with aerodynamic residuals baked into the loss.

02

CFD digital twin

An ANSYS Fluent model of near-surface airflow auto-generates a 120k-sample labeled training corpus.

03

State estimation

A temporal encoder does recursive estimation, reaching sub-3cm distance and under 4° angle error.

04

Sim-to-real

Domain randomization closes the gap so the model trained in sim holds up on real hardware.

Fig 2 — Predicted vs true distance to surface. Grey: ground truth. Red: model estimate.
<3cm
Distance error
<4°
Angle error
120K
Simulated training samples
0
Added sensors
PyTorch Physics-Informed ML ANSYS Fluent (CFD) State Estimation Sim-to-Real Time-Series UAV / Drones Python