The Best Hand in Robotics Lost to Two Fingers

Last week Google DeepMind showed off a humanoid tying a knot. The hand doing the tying belongs to Apptronik's Apollo 2, and it has five fingers and twenty-two degrees of freedom. It also sealed a ziplock bag, which is harder than it sounds and looks even harder than it is.

This week IEEE Spectrum ran an interview with a man whose robots have two fingers.

His name is Russ Tedrake. He teaches at MIT, he ran Large Behavior Models at Toyota Research Institute, and in January he took a team out of TRI to start Walden Robotics. In July they came out of stealth with about three hundred million dollars at a one-point-one billion dollar valuation, backed by Toyota, NVIDIA, Boeing, and Samsung, among others.

And since February, before any of that was public, Walden robots have been doing production work in a Toyota plant in North America. Loading and unloading car parts. Cleaning machinery. Kitting parts for assembly. They went from first pilot to real work on the floor in under two months.

Two fingers.

The Reason Is Not a Benchmark

When someone builds a simpler gripper, you expect the justification to be about cost, or speed, or a chart showing that the extra fingers didn't help on some task suite. Tedrake's reason is none of those.

"We have been deployed in a Toyota factory, and at the end of the week, the hands take a beating."

That's it. That's the whole argument. The hands come back chewed up on Friday, so build hands that can come back chewed up on Friday and keep working.

I have read a great many robotics justifications and that one has a texture the others don't. It is the sound of a person who has been in the building on Friday.

Every Other Decision Is the Same Decision

Once you notice it in the gripper, it's in everything Walden built.

Their robots have a humanoid upper body — two arms, a sensor head — on a wheeled base. No legs. The obvious read is that legs are hard, and legs are hard, but that isn't the reason he gives:

"Factories already have autonomous mobile [wheeled] robots. They already have safety cases built around AMRs."

The safety case already exists. Somebody already did the paperwork, already ran the risk assessment, already trained the floor on what to do when a wheeled thing comes around a corner. Arriving on wheels means arriving inside a category the plant has already approved. Arriving on legs means being the first of your kind in that building, which is a wonderful thing to be in a press release and a miserable thing to be in a safety meeting.

Same with the pitch. Walden doesn't claim general-purpose. Tedrake says multi-purpose — learn several genuinely valuable tasks well, get better at them with practice, don't promise the universal robot. That's a smaller claim, and smaller claims are the ones that survive contact with a purchase order.

Rodney Brooks Put a Clock on This

The same week, Rodney Brooks published an essay arguing that people constantly confuse four different time scales, which is why technology predictions come out so badly wrong.

• A new research idea to a solid lab demo: ten to twenty years. Sometimes far longer.
• Hype: months. "AI agents" went from nobody saying it in mid-2025 to everybody saying it by this month. Blockchain and the metaverse ran the same lap and left.
• Deployment at scale: twenty-plus years. Linux shipped in 1991 and Microsoft came around in 2012. Self-driving cars were demonstrated in 1987 and are still a rounding error in 2026.
• Reshaping the economy: fifty-plus years. Containerization. Electrification.

Put Walden on that chart and it clicks into place. The research clock ran at TRI for most of a decade. The hype clock is going off right now around humanoids. The clock Tedrake is actually racing is the third one — get inside real buildings, survive the week, come back Monday — and he built the entire robot around that clock instead of the other two.

And hardware moves slower on clock three than software does. Everything in physical AI is hardware.

I Have Never Installed Anything Into a New Building

Here's why I took all of this personally.

I monitor a sump pump and a forty-device building. Neither of those systems went into a new site. The pump was in that basement long before I got there. The building is a retrofit of equipment that nobody is going to replace on my account, because replacing it costs more than the building makes in a year.

I have never done a greenfield install and I don't expect to, because a customer buying new equipment is going to buy monitoring from whoever sold them the equipment. My entire business is the installed base. It's the only market at this price point.

Which means Tedrake's list is my list, and I got there the same way he did — by being cheap rather than by being smart.

There is no sensor on my sump pump. There's a smart plug between the pump and the outlet, and the system watches how much current the motor draws and for how long. Running, not running, running too long, cycling too often, all inferred from the draw. I built it that way because I didn't want to drill into somebody's pump and because a smart plug costs fifteen dollars.

But look at what that decision actually bought, in Tedrake's terms. Nothing is mounted on the equipment, so nothing on the equipment can be knocked loose, fouled, or vibrated off. The failure mode of my sensor is that a plug stops reporting, which is loud and obvious, rather than a sensor slowly drifting, which is silent and expensive. And when the pump gets replaced someday, my monitoring survives it, because it was never attached to that particular pump in the first place.

That's the two-finger gripper. I just didn't know that's what I'd built.

The Document I Should Have Been Keeping

Now the part that isn't flattering.

I can tell you in detail what my systems detect. I have accuracy stories, catch stories, the whole pitch. What I cannot hand you is a record of what has physically broken in eighteen months of running them.

I know it. It's in my head and scattered through order confirmation emails. Which plug got flaky and had to be swapped. Which sensor started reading nonsense in July heat. Which thing made me drive out on a Saturday. That history exists, it's mine, and I have never written it down.

And it's the wrong thing to be missing, because it answers the question a facilities manager is actually asking. He is not asking whether it works. Everybody's demo works. He's asking:

So that's what I'm writing this week. One table, both sites, everything: component, model, install date, cost, still running, what went wrong, when, what I did, whether I had to drive out. Including the parts that have run flawlessly for a year and a half, which is most of them and most of the value. Including the ones I replaced and the embarrassing reasons why.

It goes next to the parts list I already publish. The parts list tells you what I put in your building. The durability log tells you what happened to it afterward. One without the other is half a claim.

It takes an evening, because it's transcription rather than research. And it goes stale if I wait — the details I still remember about a failure from last summer are details I will not remember in December.

Being on the Slow Clock Is the Position

There's a version of this week where I read about a humanoid tying knots and a company raising three hundred million dollars and feel very far behind, sitting here with two sites and a smart plug.

Brooks' four clocks are a good cure for that.

The research clock on what I do is finished. Small models that catch mechanical failure run in kilobytes now, on hardware that costs less than lunch, with no cloud connection and often no added sensor at all. That question got answered, and it got answered lower and cheaper than anyone expected.

The hype clock is peaking, and it will move somewhere else, the way it always does. Nothing that happens on the hype clock will change whether a compressor in a strip mall short-cycles at 3 AM.

The deployment clock is twenty years long. It is where all the money is, and it rewards exactly one thing: showing up with something that keeps running, and having the receipts to prove it kept running.

That is a terrible position for a company that has to triple every year. It is a very good position for one person with two sites, eighteen months of history, and the ability to hand a skeptical facilities manager a document that a freshly funded competitor structurally cannot produce — because they have not been running anywhere long enough to have one.

The best hand in robotics tied a knot on video this week. The two-finger gripper was at work in a Toyota plant, where it had been since February, taking a beating and coming back Monday.

I know which one I'm building.