Ehhhhh, maybe? A fruit fly has 100k neurons. A quick search suggests that these neurons fire somewhere around 10 times per second under stress, so we get about 1 million firings per second, say.
I build audio-oriented neural networks models with ~150 MFLOPS which can run single-threaded on all kinds of phones at about 1000Hz (ie, 1000 inferences per second, each inference using about 150k multiplies). This is single-threaded on an ARM chip.
Piles of caveats: fruit flies also have to maintain breathing, etc. The neurons are more complicated than an artificial neural network activation. But they may contain a lot of vestigial crap, and an activation system which may just be a lot less efficient than silicon.
So, at least for little fruit flies, I don't know that it's THAT far off.
Can someone please explain why this is difficult? I think its trivial but I am often wrong about these things. It's got an accelerometer and a gyro. Doesn't even need vision to do this. Even simpler would be just a hard-coded program without a feedback loop except after the end of the stunt to level out the quadcopter.
Once someone starts producing a cheap drone ("missile") that can loiter and out-maneuver any human pilot if only because there's not a human on board limiting acceleration, what does the civilized world do for air superiority? It doesn't seem very many years away.
In the current Azerbaijan/Armenia war the Turkish drones own the skies consistently winning over the USSR/Russia made air defenses thus continuing the winning streak that started in Syria and Libya. (We haven't yet seen how they would do against manned planes there as Turkish F-16s have so far successfully been taking care about the Armenian planes there)
> out-maneuver any human pilot
for the last 30-40 years modern air combat usually consists of F-16/15 shooting down a USSR/Russia made plane using long range missile. I don't see any issues for the drones to carry on the tradition.
Yeah i'm in AI and i'm not sure if I should be impressed or not. It's not hard to train pretty much anything in simulation, which they did. The hardest part in robotics is the transfer to reality, which they actually did too, but i'm not sure HOW ? Is this domain randomization ? Simulation ehancement ? World model ?
FTA they created an abstraction layer that makes the abstracted simulated environment closely resemble the abstracted real world environment, which increased the transferability of the model.
> with only onboard sensing and computation