It weighs less than a paperclip. It fits in the palm of your hand. And it has just pulled off ten backflips in eleven seconds.
Researchers at MIT have presented a new control system for their flying micro-robot. The machine itself is unchanged. The software that commands it is not — and its speed has been multiplied by roughly five and a half.
Why flying small is so difficult
You might imagine a small flying robot is a scaled-down drone. It is the opposite: the smaller it is, the harder it gets.
A conventional drone is stable. It is heavy, its propellers spin fast, and the air around it behaves predictably. Gentle corrections are all that is needed.
At insect scale, none of that holds. The air becomes viscous, like honey to an object of that size. The wings flap rather than spin, creating different vortices with every beat. The machine is so light that a breath of air blows it off course. And corrections are needed hundreds of times per second.
Writing the equations for all of that is more or less impossible. This is why real insects have always flown better than our best robots.
Rather than modelling the physics, the system is left to try. Millions of flights are simulated, the machine crashes millions of times, and whatever works is kept. This is reinforcement learning, the same method that taught language models to answer correctly. Except that here, success is not measured in liked answers, but in "is the robot still in the air". The criterion is brutal and perfectly clear, which makes the learning far cleaner.
What it could be used for
The applications put forward by the researchers all revolve around the same advantage: getting where nothing else can.
Searching for survivors in a collapsed building, between slabs where a dog cannot squeeze through. Inspecting the inside of an engine or a pipe without dismantling anything. Pollinating in a greenhouse.
That last point comes up most often in this research, and it should be treated with caution: replacing bees with robots is not a solution to pollinator decline, it is a confession.
What it says about everything else
There is a point here that goes well beyond this robot.
For decades, improving a machine meant improving the machine: a better motor, a lighter material, a better-machined part. Here, the hardware stayed identical and performance was multiplied by five.
It is happening everywhere. The same camera takes better photos than it did three years ago thanks to software processing. The same car brakes better. The same prosthesis adapts better to the person wearing it.
This shift has one pleasant consequence and one less so: what you own can improve without you buying anything, and what you own depends on someone deciding to keep improving it.
The detail that remains
Eleven seconds for ten backflips looks very nice on video. But the figure that matters in this field is the one least talked about: endurance.
A robot of this size cannot carry much battery, and flapping wings consumes a great deal. As long as these machines fly for only a few minutes, they remain laboratory objects.
It is the same story as the humanoid robots that left the production line ten days ago: the demonstrations impress, and it is the boring figures that decide what makes it out of the lab.