A water tank drove the robot; the chip that would replace it has not been built

What actually drove the car

The processing layer of this setup is a water tank. The water-based experimental system, which gives direct visual access to the dynamics of the wave network, reaches near-perfect accuracy in robotic vehicle obstacle recognition, and then comes real-time autonomous motion with obstacle avoidance, controlled entirely by the wave-based hardware. An Arduino-based robotic car runs in the supplementary video.^[1]

For a medium to compute, it is enough that it offer repeatable and sufficiently rich dynamics, and a water surface supplies that for free, leaving the readout as the only trained layer. The architectural claim sits here: because both software- and hardware-based spiking neural network approaches demand substantial training and energy consumption, a simple platform is wanted. Yet no measured energy comparison against a digital controller is reported for this setup, and the advantage is attributed to the wave interactions doing the work without training. Another reading is available: if the obstacle recognition task is small enough, the advantage may come from the task rather than from the medium.^[1]

The half that is still a simulation

The path to a device is drawn through micromagnetic simulations. The concept is extended to electrically excited spin waves in a magnetic nanodevice operating at gigahertz frequencies, outlining an approach toward solid-state robotic chips based on wave computing. What stands between the water tank and that chip is fabrication: lithography, variation, temperature and yield sit between a simulated spin-wave reservoir and a fabricated one.^[1]

Which measurement would close the energy claim? Once the magnetic nanodevice operating at gigahertz frequencies leaves simulation and is fabricated, a measurement of energy consumption per obstacle recognition will be published alongside a digital neuromorphic baseline on the same obstacle recognition task, and the claim of scalable and energy-efficient robotics will be decided by that number. Until it arrives, what is in hand is an architecture that works in a water tank and its chip-scale outline.^[1]