TraviaTechPie Review

The Story

Here’s a question most people never think to ask: when a small robot keeps its balance or tracks a target, where does that “thinking” actually happen?

For another substrate-level shift — controlling light in time rather than space — see the first photonic time crystal.

The boring answer is a digital chip. A sensor reads something, that reading gets converted into numbers, a processor runs a control loop — usually a PID controller, the workhorse algorithm that’s been steering machines for a century — and then it spits out a signal to a motor. Round and round, thousands of times a second. It works. But every one of those steps costs power, time, and silicon. For a warehouse robot plugged into the grid, nobody cares. For a gram-scale drone or a tiny medical robot, that overhead is the whole ballgame.

A team in the Department of Mechanical Engineering at the University of Michigan has been chipping away at that problem, and their result is worth slowing down for. In a paper published in Science Advances in March 2025, they describe an analog control system built on what’s called “reservoir computing” — and they got it to do real robot control tasks while drawing roughly 12.5 microwatts of power. For comparison, a conventional digital PID controller doing similar work sits around 5 milliwatts. That’s about a 400-fold gap. We’re talking about the difference between a chip you’d never notice and a chip that becomes your robot’s power budget.

Let me explain what reservoir computing actually is, because the name makes it sound scarier than it is. Imagine you throw a pebble into a still pond. The ripples that spread out are a messy, complicated reaction to one simple input — and crucially, the pond “remembers” the splash for a moment as the ripples fade. Reservoir computing uses that same idea. You feed a sensor signal into a tangled physical system whose internal dynamics naturally mix and echo that signal in complex ways. The system itself does the hard, nonlinear math just by being what it is. All you have to train is a thin “readout” layer at the end that learns to pick the useful answer out of the ripples.

The Michigan team built their “pond” out of hardware. They made interconnected memristive channels — tiny components whose electrical resistance shifts depending on recent activity, giving them a built-in short-term memory — using a layered semiconductor material called bismuth selenide. To pattern those channels precisely, they used a clever trick: they rubbed the substrate to create a static-charge pattern, then let the semiconductor molecules settle only into the charged regions. Don’t worry about memorizing that. The point is, the resulting network behaves like a physical reservoir. Sensor signals go in, get scrambled into a rich high-dimensional swirl, and a simple trained layer reads out a motor command.

And it works on real tasks. They demonstrated two: a rover tracking a moving target, and a drone-style motor balancing a lever. In both cases the analog system matched the behavior of a traditional digital controller — same job, wildly less power, far less silicon, and almost none of the analog-to-digital conversion overhead that normally eats into a robot’s energy budget.

That last part is the quiet headline. The big savings here isn’t just a more efficient processor. It’s that the computation barely happens as “computation” at all. The physics of the material is doing the work. There’s no fat software stack, no constant shuttling of numbers between sensor and CPU. The control loop is closer to a reflex than a calculation.

The Takeaway

Here’s why this matters more than a 400x efficiency number suggests.

This blog has been tracking a theme we’ve called “Physical AI” for a while now — the shift from AI that lives in a data center to AI that has to move, balance, and act in the messy physical world. When we covered the Physical AI investment wave back in February, and Tesla’s Optimus, and humanoid platforms like LeRobot, the conversation was mostly about bodies and brains: better actuators, better policies, more training data. What gets discussed far less is the substrate — the actual hardware the control runs on. And that substrate is becoming the bottleneck.

Here’s the thing. A humanoid robot or a swarm of insect-scale drones doesn’t fail because its AI model isn’t smart enough. It fails because smart costs watts, and watts cost battery, and battery costs weight, and weight kills the very agility you were trying to build. Every milliwatt you spend on a digital control loop is a milliwatt you can’t spend on flying longer or carrying more. This is exactly the corner that reservoir computing sidesteps — by letting the hardware itself be the computer.

It also rhymes with research we’ve covered before. When we wrote about USC’s high-efficiency artificial neuron, and Georgia Tech’s soft robotic “eye” that runs with no external power at all, the through-line was the same: stop treating computation as a separate digital box bolted onto the robot, and start baking intelligence into the materials and physics. The Michigan work is another data point in that pattern. My read is that we’re watching the slow birth of a third layer in robotics — not just hardware and software, but a “physics layer” where the line between the two gets deliberately blurred.

Now, the honest caveats. This is a lab demonstration, not a product. Two control tasks is a proof of concept, not a platform. Analog systems are famously twitchy — they drift with temperature, they’re hard to manufacture identically at scale, and a memristive network that needs static-charge patterning is not something you stamp out by the million next quarter. Reservoir computing also has a real ceiling: it’s brilliant for low-level, reflex-style control, but it won’t be planning a path across a warehouse or reasoning about a task. That’s still digital territory.

So the realistic future isn’t analog replacing digital. It’s a split. The heavy cognitive work — perception, planning, language — stays on power-hungry digital chips, probably offloaded or batched. But the fast, boring, always-on reflex layer — balance, stabilization, motor trim — drops down onto microwatt analog hardware that just reacts. Your body works roughly the same way: you don’t consciously think about staying upright, a faster and cheaper system handles it. Robots have been doing it the expensive way because that’s all we had.

Here’s the point worth holding onto: the next leap in robotics may not come from a smarter model. It may come from a chip that finally stops calculating so hard — and lets physics carry the load.

Photo: Umberto / Unsplash