Researchers have trained robot control software small enough to fit in a few kilobytes of memory - and it still works on real hardware.
Researchers used evolution strategies paired with statistical model checking to search directly for compact neural network policies, rather than training a large network and compressing it afterward. They tested the method on two classic control problems, a cartpole balancer and a quadrotor, across a range of control frequencies and network sizes. The resulting policies take up between 0.5 and 7.5 kilobytes of memory - small enough for a basic microcontroller - and transferred from simulation to physical hardware with no extra retraining. On that hardware, the policies ran in real time with under 25 nanoseconds of timing jitter, while the chip sat idle more than 97 percent of the time.
Most robotics AI assumes a GPU or a hefty CPU is doing the thinking nearby. This work assumes the opposite: a control policy that fits in a few kilobytes and leaves a chip almost entirely free can run directly on cheap, tiny, battery-constrained hardware, without offloading decisions to a bigger onboard computer or the cloud. The statistical model checking step matters too - it's a formal check on safety and robustness built into the search, not bolted on after the network got small.
It's a long way from balancing a simulated pole to flying a real drone through a warehouse, but if those safety guarantees hold up outside the lab, memory and power budgets - not raw model size - start looking like the real limit on where robots can go next.