Sir Airffic GoMuch of this comes from research papers, but the words are kept simple. Stay with me.
Four stages of understanding
- Detection: find people, cars and power lines in an image and draw boxes around them.
- Tracking: follow where a found object moves next.
- Self-localisation: work out how far and where you moved from image motion plus the IMU (VIO: visual-inertial odometry).
- Understanding and instruction: grasp "check behind the house with the red roof" and link the words to what you see (VLM: AI that handles images and language together).
Flying while mapping: SLAM and VIO
In the late 2000s, university labs began estimating a small drone's position with one camera and an IMU. Everything started there.
Methods such as SVO, VINS-Mono and ORB-SLAM were published and became shared foundations. Open research turned into product components.
Today's commercial drones can fly under bridges and indoors because of this accumulated work.
From above, things look small and skewed
AI trained on ground-level photos struggles with views from above: people become dots and cars change shape.
Aerial datasets such as VisDrone and DOTA were built so that sky-specific AI could be trained.
Whoever owns the data shapes how smart the AI gets. That has not changed in the foundation-model era.
Sir Airffic GoAI gaps often come down to how much aerial data you own, more than to algorithms. Remember this for the strategy chapter.
The day AI beat the human champions
In 2023 an AI drone called Swift, built at the University of Zurich, beat human world-champion drone racers. The work was published in Nature.
Swift estimated its position with only a camera and an IMU, and learned to fly through reinforcement learning — practising many times in simulation.
This was on a fixed course. Matching that skill in the real sky, with changing weather and obstacles, is still an open challenge.
Champion-level drone racing using deep reinforcement learning (UZH)
Evidence: [N124]
The next wave: see, reason and act on words
From around 2023, large AI models that handle images and language together (foundation models) began to serve as robot brains.
Drone research on following spoken instructions while understanding the scene (aerial vision-language navigation) is growing.
Large models are heavy, and running them fast on board is still hard. That is why chips — the next chapter — must be read alongside.
Evidence: [N133]
Frequently asked questions
What jobs use drone computer vision?
Crack detection on bridges and power lines, crop health checks, finding people in search and rescue, security and traffic monitoring, and checking delivery sites are common uses.
What is visual-inertial odometry (VIO)?
A method that combines image motion with the IMU's measurements of acceleration and rotation to calculate how far the drone moved. It is a basic technology for flying without GPS.