How Small Can a Neural Network Be and Still Control a Robot Safely?

The networks that fit in a kilobyte

Micro Neural Policies are not compressed versions of larger models. They are synthesized from scratch to stay tiny. The memory footprint ranges from 0.5 to 7.5 kB. That is smaller than a single high-resolution photograph. A typical microcontroller can hold dozens of them at once.

The authors are Hongpeng Cao, Riccardo Curcio, Daniele Ottaviano, and Marco Caccamo. Submitted 6 Oct 2026. Nine pages.

How evolution and verification shrink the policy

Evolution Strategy searches the space of possible networks. Statistical Model Checking verifies each candidate against safety properties. The loop runs until the policy is both small and provably robust. No gradient descent. No massive datasets. The method works on Cartpole and Quadrotor tasks. Control frequencies and architectures vary across the experiments.

From simulation to a physical quadrotor

Policies train in simulation. Then they transfer to hardware without fine-tuning. The paper calls this zero-shot sim-to-real transfer. Control performance does not degrade. Safety guarantees hold. The quadrotor flies. The cartpole balances. The jitter stays under 25 ns. That number matters for hard real-time loops.

Over 97% of the processor time remains free.

What the chip does with the spare cycles

The inference itself takes a fraction of a microsecond. The rest is available for communication, logging, or a second control loop. The paper does not say how many such policies have been deployed in products. It demonstrates feasibility on representative benchmarks.

Where the explanation stops

The work shows a path. It does not claim the path is finished. Larger robots, richer sensors, and adversarial environments are not covered. The verification scales with state-space complexity in ways the paper does not quantify. The next step belongs to engineers who need to decide if 7.5 kB buys enough intelligence for their specific machine.