In the process of continuing to explore the extreme space flight of UAV, the joint team of the University of Lorraine in France and Inria recently released its phased results. In June 2025, they published an article in the Nature sub-journal 《npj Robotics》 Publish a paper Flying in Air Ducts, proposed a lightweight flight system that combines aerodynamic sensing and neural network localization, and successfully realized an 18-centimeter micro-quadrotor flying with a diameter of only 35cm of in ventilation duct autonomous flight. The system integrates aerodynamic modeling in the air duct, ToF-IMU fusion perception and neural network localization, breaking through the key bottleneck of autonomous hovering and navigation of micro UAVs in narrow spaces.

Video source: __https://www.youtube.com/watch\?v=7je2hUnGwms

01 Research background

Ventilation ducts are key structures in modern buildings (such as commercial buildings, hospitals, rail transit, and industrial plants). Their internal conditions directly affect the operation of air conditioning and heating systems and air quality, and require regular inspections. However, these closed, narrow, and complex air duct systems pose great challenges to traditional detection methods:

• Manual work is not allowed;

• It is difficult for ground robots to climb/overcome obstacles;

• The aircraft faces severe aerodynamic interference and perception obstacles.

Although some wheeled and shelled flying robots have been used in pipeline environments, they are mostly used in larger spaces or rely on manual remote control, making it difficult to achieve stable autonomous flight in small-diameter circular air ducts.

Technical difficulties

• Aerodynamic disturbance is complex The backflow inside the air duct is strong, and the superimposed ground effect, ceiling effect and wall effect make the UAV highly sensitive to hovering and prone to instability.

• Perceptual conditions are extreme The air duct environment is closed, dark, and featureless, making it difficult for visual localization algorithms to work, and the traditional ToF+IMU method lacks accuracy.

• Computing resources are limited The computing power of embedded platforms is limited and it is difficult to run high-complexity algorithms in real time. It is necessary to implement low-latency, highly robust, and deployable localization and control solutions.

02 research methods

This research proposes a set of methods for autonomous and stable flight of micro UAVs in narrow ventilation ducts.Fusion of aerodynamic modeling and end-to-end position estimation The lightweight flight controller system achieves hovering and autonomous flight in an air duct with a diameter of only 35 cm.

Image source: "Flying in air ducts", Thomas Martin et al., npj Robotics 2025

Aerodynamic field modeling in ventilation ducts

The research team first quantified the return aerodynamic force of the rotor in the closed circular tube and determined the most suitable position for hovering based on this.

• The research team fixed the UAV model equipped with a six-dimensional force/torque sensor at the end of the Franka Panda seven-axis robotic arm, and measured the aerodynamic force difference statically at 192 discrete points defined by the circular tube section.

• Based on the measurement results, the system drew the disturbance force distribution diagram of the circular duct cross-section, revealing that the rotor return flow superimposed on the wall adsorption effect resulted in a non-uniform distribution of aerodynamic force with height and lateral position.

• The data shows that the distance from the bottom of the tube is about 10cm The additional force is minimal at the hovering height, which is a relatively stable height; the disturbance at the center line and near the top increases significantly.

Image source: "Flying in air ducts", Thomas Martin et al., npj Robotics 2025

Data-driven lightweight localization system

In order to solve the problem of visual failure caused by textureless and weak light in the pipeline, the paper proposes a localization solution of ToF sensor array + embedded MLP.

• Micro ToF sensor array: Use 9-10 VL53L1X sensors, pointing in different directions around the UAV, to measure the distance between the UAV and the pipe wall. The sensor has high ranging accuracy, fast response, and is not affected by the light environment.

• Neural network architecture: The research team built a multi-layer perceptron (MLP) model that inputs ToF sensor data, UAV speed information, and IMU attitude information to output the UAV's lateral and longitudinal positions within the pipeline section in real time.

Image source: "Flying in air ducts", Thomas Martin et al., npj Robotics 2025

• Model training: Using OptiTrack and HTC Vive Lighthouse system, 52 minutes of precise flight data were collected, and a training set and a test set were constructed. The neural network finally achieved localization error within 1 cm Within.

• Edge computing deployment: After optimization, the neural network can be directly deployed on the STM32 microcontroller mounted on the UAV, with strong real-time performance and a delay of less than 1 millisecond.

03 Experimental testing

The research team used a The whole machine is about 18 cm wide (propeller tip to tip) and 7.5 cm high. The total weight including battery is about 130g. Micro quadcopter for flight testing.

• At the optimal height (~10cm), the UAV hovers stably for more than 2 minutes in a circular tube with an inner diameter of 35cm;

• In pipes with inner diameters of 45cm and 56cm, the neural network model shows good generalization ability;

• In a tunnel with a total length of 3.9m and an inner diameter of 45cm, the 1.5m forward and return test (3m in total) was performed, and the lateral error was <2cm.

Image source: "Flying in air ducts", Thomas Martin et al., npj Robotics 2025

Resource Express
Paper link: https://www.nature.com/articles/s44182-025-00032-5