Build log
How our robot chooses where to land.
A shelf, a table, a stool: a small flying robot could land on many things in a home. The hard part is knowing which ones are good enough, from a few readings, while hovering. Ours is designed to land only where it is sure, and to look elsewhere when it is not.
What it checks before it lands
It flies to a point above the spot and looks down with its distance sensor for a moment, taking many readings in a row. From those it works out where the surface is and how it tilts, and it lands only if every check passes. The whole landing pad has to rest on the surface with room to spare. The surface has to be level enough. Nothing may stand where the propeller guards will be, no edge or drop may be too close, and no person may be within a metre. It does not land on glass. Then it comes down slowly, watching the edges as it goes, and backs off if anything changes.
Thousands of rooms that do not exist
We test this in simulation, in rooms built at random: surfaces of many sizes and heights, in wood, dark material or glass, with things standing on them and a lamp, a wall or a person nearby. A separate check that knows the true geometry decides whether landing was possible at all, so we can tell a wise refusal from a missed chance.
What a wall taught us
In early runs, the robot touched a wall a few times on its way to a landing, always before the landing itself. Tracing those flights showed one pattern: the route climbed over a wall and then came down right at its top edge, while the robot’s idea of where it was had drifted by a few centimetres.
We added a rule. On the way to a landing, the robot does not descend past the top of anything within its reach until it is clear of it. Since then, those flights end in a refusal instead of a touch, and it lands as often as before.
Careful first, then more often
Being strict has a price: the robot passes up some places where it could have landed. We are working on that price, not on the caution. A small network now suggests where to look next when the rules find nothing (what our ring and robot learn), and better position estimates should let the rules accept more. When nothing suitable is in reach, it says so instead of guessing. Moving air is still the hard part, and that is where our flight controller work is now: what a fan does to a small drone.
What this is not
These are simulation results, with sensor noise and position errors set from published values rather than measured. The rooms are random, not real homes, and real homes have cushions, plants and cats. Alkyon is a concept that flies only in simulation.