Products — Ülgen Aero

How a mission runs

Define what will be inspected, let the drone fly it, review what it found and share the report.

1
Plan
Define the structure, the risk boundaries and the facade zones to inspect.
2
Fly
The drone follows the planned route and captures close-range imagery.
3
Analyse
Imagery is reviewed so defects and anomalies stand out.
4
Report
Findings are delivered as a clear report the team can act on.

Interactive 3D model

Meet the drone

The drone is flown by a reinforcement-learning flight policy trained in NVIDIA Isaac Sim. It reads the drone’s current state and chooses actions that are applied as thrust and torque on three axes. During training, rewards guide the policy towards a steady path close to the facade. While it flies, a separate detection model analyses the camera stream to identify both cracks and surface materials. Engineers review the resulting findings.

Loading the interactive 3D model...

Drag to rotate. Pinch or scroll to zoom.

Trained in simulation

Before it flies near a real wall, the flight policy learns in Isaac Sim. Many drones practise at once, among building blocks and along facades, and every flight feeds the next round of training.

Isaac Sim, where the flight policy is trained.

Many drones train in parallel.

Simulated city of white building blocks on a dark grid — Building blocks for obstacle training.

Simulated drone flying close to a white building wall — Flying close to a facade.

Navigating between blocks, top-down.

Simulation in action 1 / 5

Simulation environment

Control loop

State observations enter the flight policy.

01Flight policyChoose an action from the current observation.
02Apply actionThrust and torque on three axes.
03Physics stepPhysX updates the drone state in Isaac Sim.

Repeat with the new state observation.

Training only: rewards and collected experience guide PPO updates to the policy. Rewards are not required as flight-policy inputs during deployment.

Perception pipeline

Camera imagery enters the detection model.

01Drone cameraCapture facade images during inspection.
02Detection modelIdentify both cracks and surface materials.
03Findings and reportCombine detections and material context for engineer review.

One flight model. One detection model.

The inspection pipeline is separate from the PPO flight-training loop.

On-site AI review

Defect detection

One AI model controls the flight. A separate detection model identifies both cracks and surface materials in the camera images. Material context helps engineers interpret a crack in concrete, brick or stone and assess the need for follow-up.

Illustrative drone camera view of a cracked concrete building facade — 1Camera frame The drone captures facade imagery. Illustrative camera view.

Illustrative cracked concrete surface — 2Material detection The same detection model identifies the surface material. Illustrative surface examples.

Illustrative cracked brick surface — 2Material detection The same detection model identifies the surface material. Illustrative surface examples.

Concrete wall with a crack outlined by a detection box labelled crack, confidence 0.67 — 3Crack detection Each crack is outlined and given a confidence score.

4Assessment and report Findings and material context are organised for engineer review.

Where it works

The same workflow adapts to the site and the surface being examined.

Buildings
Facades, exterior walls and other elevated surfaces where access is slow or disruptive.
Bridges
Large structural surfaces and hard-to-reach sections that need repeatable coverage.
Dams and concrete structures
Broad surfaces where visual records and organised follow-up improve the inspection.