Before you step off a curb, you do something so automatic you barely notice it: you look both ways. You turn your head, sweep your eyes across the road, and only then commit to crossing. It is a tiny pause, but it is the difference between walking and walking safely. A drone that flies indoors needs the same instinct — and this is the story of how a small quadcopter learned to scan its surroundings on purpose before moving, and why that deliberate sweep makes it a fundamentally better explorer than one that flies blind.
The difference between seeing and looking
Our drone carries two very different kinds of sensors. Around its body sits a ring of short-range laser sensors — cheap, always-on, looking outward in every direction at once. They are the equivalent of your peripheral vision and your sense of touch: great for “there is a wall right next to me, don’t clip it,” useless for anything further out.
The interesting sensor is a single long-range rangefinder, a TF-Luna, mounted on a tiny rotating neck. It can see far — several meters down a corridor — but only in the exact direction it happens to be pointed at that instant. Everything outside that narrow beam is, as far as the drone is concerned, darkness. To know what is down the hall, the drone has to physically turn the beam that way and look.
That is the crux of active perception: the drone does not passively receive a complete picture of the world. It has to act to perceive. Just as you cannot see behind you without turning your head, the drone cannot know what is ahead without aiming its scanner there first. Perception becomes a choice, not a gift.
Why a cheap rangefinder forces the issue
Why not just bolt on a fancier sensor that sees everything at once? Because the whole point is to do good indoor mapping with hardware that is genuinely affordable. A single-beam rangefinder costs almost nothing. The catch is that its narrow view turns a hardware question into a behavior question: the drone’s effectiveness depends not on what it can see, but on where it chooses to look — and when. Aim the beam well and it builds a rich, current map of the room. Let it wander and the drone is mapping by luck. With a cheap sensor, looking is a skill, and the skill is everything.
From a hardcoded sweep to a learned habit
The obvious approach is to make the scanner spin automatically, like a lighthouse. We tried living with that, and it failed in a subtle way: the beam swung around, but aimlessly. The drone leaned entirely on its short-range ring and groped through rooms by feel, ignoring the long-range tool meant to let it look ahead. The expensive capability was wasted because nothing made the drone need it. Worse, in small rooms the always-on short-range ring was simply enough to avoid bumping into things, so the scanner was redundant — and the drone sensibly ignored a tool it did not need. A hardcoded sweep does not teach a drone to care.
So we changed the rules instead of writing a routine. First, we shrank the short-range ring down to a last-ditch collision bumper — enough to say “don’t hit the wall in front of you,” but too short to navigate with. To know what is further ahead, the drone now genuinely has to turn the scanner and look. Second, we made flying into unscanned space costly: move into a region you have not looked at recently and it is treated as unknown territory, and pushing into the unknown is discouraged.
We trained the policy with PPO — think of it as a coach that watches thousands of attempts and nudges the drone’s instincts toward what works. Crucially, we never told it how to scan. We made blind flight expensive and let the coach rediscover the habit on its own. Looking before leaping was not programmed; it emerged, because it was the cheapest way to succeed.
What the drone learned to do
The result was clean. The drone kept exploring every test room without crashing — even though, up close, it can now only see half a meter. It had genuinely learned to lean on the scanner. Blind flight, where it moves into unscanned space, dropped dramatically: from roughly a third of its moves down to a small fraction. And the scanner stopped wandering — it now brushes across the whole sector ahead, like a little radar sweep, before the drone commits to a move.
Interestingly, the drone did not learn to point the beam precisely along its flight path. It learned to sweep the area ahead quickly and then advance — a broad “looked around, then went” strategy rather than a narrow stare. That is still exactly the instinct we wanted: look, then go.
This builds on earlier work that made the drone’s flight honest — giving it real inertia, real tilt, and a scanner the agent steers on purpose rather than one the simulator spins for free. Once the drone could no longer stop on a dime or see for free, looking ahead stopped being optional.
Design takeaways
- Make the right behavior necessary, not mandatory. A penalty for blind flight beat any hardcoded “scan now” rule. Remove the crutch and the skill appears on its own.
- Active perception is a choice the agent must make. With a narrow-beam sensor, where you look is part of the policy, not a fixed routine.
- Cheap hardware shifts the burden to behavior. A single-beam rangefinder can map a room well — if the drone learns to aim it well.
- Let the learner discover the “how.” We defined the goal (don’t fly blind) and let PPO find the strategy (sweep, then advance).
- Honest physics first. Looking before you leap only matters once you can no longer stop instantly or see everywhere at once.
A drone that pauses to scan sees more than one that flies blind — the same way a pedestrian who glances both ways crosses more safely than one who steps off the curb without looking. The trick was not building the reflex. It was building a world where the reflex was worth having.
Further reading: dev-log 44 — honest flight dynamics and the first conscious scanner. dev-log 52 — turning scanning from a quirk into a skill.