Scientists Taught Drones to Fly Themselves With Simpler AI Eyes

Most drones that fly themselves today rely on either hand-written rules or AI that watches full video feeds. Both have problems. Hand-written rules break when the wind picks up or the target moves oddly. Full-video AI is expensive to run and famously unstable to train — the drone crashes a lot before it learns anything useful. A new paper from researchers Saurbh Singh Jamwal and Nived Chebrolu takes a middle path, and the results are surprisingly good.

Instead of feeding the drone a raw camera image, the researchers use a lightweight system that picks out just the target and reports a few key facts: how far left or right the target sits in the frame, and roughly how far away it is. Those get combined with basic sensor readings — how fast the drone is moving, which way is down — into a compact 12-number summary. Think of it like navigating by a few dashboard gauges instead of watching a whole movie. Less to process, less to go wrong.

They then trained the drone's AI in stages, like teaching a student with gradually harder homework. First the target appears in easy spots. Then the drone is allowed more aggressive maneuvers. A third version combines both. All three learned to fly well, but the staged approach — especially the version that ramps up visual difficulty — was the clear winner when the target started moving in ways the drone had never seen before.

One finding surprised even the authors: the drone did not actually need explicit distance measurements to fly well. What it truly needed was its own body-sense (speed, orientation) for stability, plus simple left-right framing to aim. Against traditional tuned controllers, the AI pilots were notably tougher under strong disturbances. Real-world payoff: cheaper sensors, more stable drones, and delivery or inspection aircraft that cope better with unpredictable conditions.