The Cost of Watching Humans
Why safety is the energy bottleneck of humanoid robots — and how we could ease it
Intro
From Palo Alto, the robot invasion is no longer a prediction. It is traffic. The white Waymos with the rooftop dome are part of the ordinary flow — 3,000 vehicles, 500,000 rides per week, the entire Peninsula covered from San Francisco to my home. They are no longer a novelty that makes headlines. They are driverless Ubers. You casually take one to dinner.
Thirty miles from here, on May 9, 2026, the last Model S rolled off the Fremont production line. Fourteen years, ending on a Saturday. Not to make room for a new car. To make room for Optimus.
According to Tesla’s earnings call, Fremont’s S/X line will be converted into an Optimus production facility, with a long-term target of 1 million units per year.
After all, Musk has always said it: Tesla is a robot. It moves like a robot, it has the autonomy of a robot. Now it will have legs instead of wheels.
This is yet another new normal: cars drive themselves, at scale. Car factories become robot factories. Robot BOMs are the real battleground between hyperscalers and nations. The success of a product increasingly depends on choices such as LiDAR versus cameras, or the hybrid boundary between edge and cloud.
And the most important question we should be asking about humanoid robots is this: how will we keep them close to human beings without consuming half their battery just to watch us?
1. Thirty Thousand or (Perhaps) One Million
Tesla announces 1 million Optimus units per year and produces a few hundred. Hyundai announces 25,000 Atlas robots and already has the actuator factory. These are two different theses applied to the same object.
Let us stay on the conservative side.
On May 20, 2026, just one week ago, Hyundai Motor Group presented at JPMorgan in Boston a plan to deploy 25,000 Atlas robots across its American factories by 2028.
30,000 units per year at full scale. Hyundai is not Musk. It promises numbers built on a supply chain that is already planned and customers that are already guaranteed — 83% of produced units are destined for Hyundai and Kia.
And then there is China.
In 2025, roughly 17,000 humanoid robots were shipped worldwide. 14,400 came from Chinese factories. 85%. 140 active manufacturers distributed across industrial clusters from Wuhan to Shenzhen.
The Chinese government is fully aware of the ongoing robot invasion and has launched the Humanoid Full Lifecycle Management Service Platform. Every humanoid robot built in China receives a 29-character identification code. Model, manufacturer, declared intelligence level, training history, maintenance logs, battery status, joint wear. The system is modeled after the national identification system used for Chinese citizens, with 11 additional characters to cover operational machine data. 28,000 robots have already received an ID.
The market-access rule is explicit: no code, no market.
We are not yet at the trajectory-based identity system I discussed in I Am More Than the Sum of the Serial Numbers of All My Parts, but it is a remarkably good start.
Everyone wants to move robots out of factories and into spaces that were never designed for them. Streets, sidewalks, mixed warehouses, hospitals, homes.
And this is where a completely new set of problems and paradoxes begins.
2. The Physical Paradox
A car moves. A domestic robot hardly ever does.
It sounds trivial, yet it is the key to understanding everything else.
A commuter Tesla moves roughly 4,000 pounds over 60 miles per day — commuting from home to work and back in an urban environment. 15 kilowatt-hours dedicated to propulsion alone. On top of that comes autonomous-driving perception: cameras, radar, compute. Somewhere between 10% and 30% of instantaneous power. It costs roughly 25% of the vehicle’s range, and it is a cost you accepted without even noticing when you bought the car. Tolerable, because it is spread across an ocean of motion energy.
A domestic humanoid weighs roughly 130 pounds. Inside a 1,600-square-foot home, even assuming 200 room-to-room movements per day — a pace that no human butler would sustain — it might travel about 1 mile.
130 pounds multiplied by 1 mile: 130 pound-miles per day.
240,000 for the car.
130 for the robot.
The ratio is roughly 1,800 to 1.
Even under very generous assumptions for the robot — three times the movements, a larger house — the ratio never drops below 500 to 1. Three orders of magnitude of difference in mechanical work between the car you use to commute and the robot you wish you had in your kitchen.
There is a second number that tells the same story from another angle.
The car is continuously moving during its mission: roughly 1 hour of near-continuous driving. The domestic robot, with its mile of actual walking spread over 16 hours of operation, is moving only 2% or 3% of the time. The remaining 97% is spent standing, waiting, making micro-movements. Setting the table. Waiting. Manipulating objects slowly. Supervising.
The car is a machine-for-movement.
The domestic robot is a machine-for-presence.
These are two different physics models, and that changes everything.
For both of them, however, the safety sensing system never turns off. It cannot. Ever.
A 2.3-kilowatt-hour battery — Figure F.03 (official: Figure AI, F.03 Battery Development) and Tesla Optimus (widely reported, but not officially confirmed by Tesla) — lasts between 2 and 5 hours of dynamic operation.
That translates into an average power draw between 400 and 1,200 watts, depending on how hard the robot is actually working its joints. Perception and compute consume between 50 and 160 watts continuously. A solid-state LiDAR is frugal — around 3.5 watts. The real problem is the brain: the Jetson AGX Thor running continuous sensor fusion and visual inference, consuming between 40 and 130 watts. That is the expensive part, because interpreting the world costs more than measuring it.
On a moving car, against 9,000 watts of propulsion power, those 150 watts are background noise. They almost disappear.
On a walking robot, against roughly 900 watts of average dynamic locomotion, they represent around 10% to 15%. Still manageable.
On a domestic robot loading dishes into a dishwasher, standing still with its base stationary — roughly 350 watts of average static power consumption — those same 150 watts become up to 47% of the total energy budget. The robot burns up to half of its battery watching the environment while barely moving.
This is the cost of watching humans.
The ratio is exactly the opposite of the car.
The car justifies sensing because it moves. The domestic robot pays for sensing precisely because it does not move. The more stationary it becomes, the worse the tradeoff gets. The cost of safely existing does not scale down the way movement does.
Increasing battery capacity does not solve the problem. Doubling kilowatt-hours means doubling weight, volume, cost, and charging time. It also triggers the inverse calculation: more battery means more mass to move when the robot walks, therefore more propulsion energy, therefore more battery. It is a loop that never closes.
The problem is not battery capacity.
It is the architecture of safety.
3. Feeling Water
A human body does not wait for the brain to finish interpreting a scene before reacting to something approaching too quickly. It uses reflexes. Peripheral layers. Local signals that trigger a motor response before the visual cortex has finished processing what was about to hit you. The brain understands afterward. The body reacts first.
Humanoid robots today are being built almost exclusively as cognitive machines. Everything goes through the brain. Nothing through the reflex.
And yet a robot does not need to see you to know you are there.
A human body is 60% water. A conductive column roughly human-sized standing in front of a machine. From the perspective of an electromagnetic field, you are a very specific perturbation. Wood does not create it. Plastic does not create it. A chair does not create it. You do, and you do it in a way that no other object inside a home replicates. You have a unique signature.
From that moment on, proximity safety stops being a vision problem.
It becomes a measurement problem.
A capacitive sensor emits a weak electric field. When you approach, the disturbance is registered as a variation in capacitance. It is a number.
A continuous stream of numbers entering a microcontroller.
No images. No frames. No sensor fusion. No visual inference model.
No sensitive data. No cloud. No network latency to depend on before deciding whether to slow down an arm. Fractions of a watt.
At that point, comparing it with a visual system almost hurts.
A camera produces 10 to 30 megabytes per second of raw data. Every frame is compressed and sent to onboard compute — a Jetson AGX Thor consuming 60 to 130 watts — where a vision model attempts to recognize people, postures, and distances. Tens of thousands of active parameters.
Billions of operations per second. All to arrive at a decision that is ultimately binary: is somebody too close, yes or no?
The numerical stream produced by a capacitive sensor answers the same question using a calculation that runs on a microcontroller costing only a few dollars. 15 milliseconds total, from field perturbation to mechanical actuation — 7 milliseconds to measure and decide, 8 milliseconds to react. During those same 15 milliseconds, a 60-frame-per-second camera produces a single frame. The visual system still has to interpret it.
No training. No connection. No brain.
The BOM changes by an order of magnitude. Power consumption changes by an order of magnitude. Latency changes by an order of magnitude. Even the risk category changes: there is no model that can hallucinate a wrong identification. There is a field. You perturb it or you do not.
This is the reflex layer that domestic robots are missing today.
There is a detail that changes everything once you see it. The visual stack used by humanoid robots today is inherited from autonomous vehicles. It was designed to solve a different problem — recognizing objects at 60 miles per hour, 100 yards away, under variable lighting conditions. It was transplanted into robots because it already existed. But it remains a system designed for high-speed motion, now applied to a machine that spends most of its life standing still inside a small room while people walk 8 inches from its torso.
It is oversized. It is expensive. It consumes power. And for proximity safety — the thing that truly matters when a 130-pound robot is standing next to an elderly person — it may not even be the right technology.
Author's note: Readers of this Substack know that I live in Palo Alto but spend much of my time working with manufacturing companies in Italy and Germany. That is where I first saw the technology I am describing implemented in practice. It was invented by a startup spun out of Scuola Superiore Sant’Anna that I encountered through Innovit. Marcello, its founder, built the entire company around freeing industrial robots from cages. I saw something else: a way to bring the domestic butler one step closer. A solution capable of balancing energy consumption and human-interaction frequency. Today, AuraSensae — that is the company’s name — also has an office in Palo Alto, next to Tesla Engineering headquarter.
One thing must be said clearly. In its basic form, capacitive sensing operates across only about a foot. The robotic-collaboration configurations does not replace cameras. It does not know where you are in a room. It only knows whether you are approaching the robot’s torso or arm. It is a shell. It works on the body of the machine, not on the environment. And it works because those few inches around the robot are exactly where the decision about human safety is made.
The visual system for the world.
Capacitive sensing for the body.
Cognition to understand.
Reflexes to react.
And within those few inches: no brain, no cloud, no inference. Just water disturbing a field, and a number that changes.
4. Conclusion
The domestic butler caring for our elderly parents will not arrive in 2027. Nor in 2030.
The robots that Figure AI already shows working on BMW assembly lines and elsewhere are physically tethered. Literally. Connected to a wall outlet through a cable.
Hyundai’s Atlas at Metaplant Georgia will also work tethered, as if carrying an umbilical cord that openly reveals how young this creation still is. Just like every industrial robot and collaborative robot before it, whether protected by a cage or not.
Without any ambition to turn this into a scientific publication, I attempted to visualize this relationship between energy consumption and frequency of human interaction through a chart that helps clarify where the various use cases sit. Here it is, even if only in alpha form.
We will not build the domestic butler by increasing battery capacity. We will not build it by adding more cameras, more LiDARs, or more Jetsons inside the robot’s chest. That path leads to a machine that burns half its energy budget standing in a living room watching you.
We will build it by removing work from the brain.
By shifting proximity safety from visual perception to the reflex layer.
By letting the water in your body perturb a field and change a number instead of asking a 130-watt model to determine who you are.
The trajectory is already visible in the chart.
It begins with the tethered factory robot.
It moves through the cage-free collaborative robot.
Then the mobile robot in structured environments.
And eventually — one day — reaches the domestic butler that can care for your mother without exhausting its battery before lunchtime.
Every step is achieved by solving one piece of the BOM at a time — piece by piece, sensor by sensor, watt by watt.
Waiting for Daneel R Olivaw .... might take a while