Jonathan Scholz spent seven years running the applied robotics team he founded at Google DeepMind in London. The company he started after that came out of stealth on Monday, and its pitch has almost nothing to do with model size. Reimagine Robotics wants the person who already does the job to teach the robot by showing it. Scholz calls the approach monkey-see, monkey-do.
That sounds like an AI claim. It's a product decision, and it's the one robotics founders are most likely to make by accident.
According to The Robot Report, in one deployment at an electronics disassembly facility, where a three-robot cell pulls valuable material out of used hard drives, Reimagine says prototyping and testing a new robot behavior went from about a day to about ten minutes. A worker demonstrates the task, watches the robot try it, and corrects it while it runs. No specialist programmer in the loop. Another customer, a made-to-order plastics shop, has the robots tending 3D printers overnight: removing beds, working latches, pressing controls. Scholz founded the company in April 2025 with three former DeepMind colleagues, and raised a pre-seed round from Fly Ventures and firstminute capital. The amount has not been disclosed.
Two answers to the same question
Three weeks earlier, Walden Robotics launched at a $1.1 billion valuation with $300 million behind it. Russ Tedrake, the MIT professor who ran large behavior models at Toyota Research Institute, is co-founder and CEO. Toyota led the round. NVIDIA and Boeing are in it. Walden's semi-humanoid robots, a torso with two arms on a wheeled base, have been working inside a Toyota plant in North America since February, moving from pilot to production manufacturing and logistics tasks in two months. Tedrake says the company is building the "full stack": hardware, software, AI, and the application layer that sits on top.
Both companies name machine tending as a target task. Both are selling robots that learn on the job. What separates them is who does the teaching once the machine is standing on somebody else's floor. Walden is building that layer itself. Reimagine is handing it to the customer's own operator.
I don't know which one wins. I do know most founders never write the decision down.
The spec is not the job
Years ago I sold a terminal emulator to government contractors. One of the terminal types we had to support was the 3179G, and the spec on it ran 1,800 pages. Implementing the whole thing would have taken years, and we were planning the work like it would.
Then we found out what the customers were doing with those terminals. They wanted to display CAD files so they could verify parts. That was the job. It was that simple. The work collapsed into a six-month coding project, and we still sold tens of thousands of seats for millions of dollars.
The 1,800 pages described everything the terminal could do. The customer needed one thing it did, and nobody on my team could have found that gap by reading the spec, because it only existed in the heads of the people doing the work.
That gap sits inside every general-purpose robot on the market. Machine tending is not one task. It is a particular latch on a particular machine in a particular cell, plus a knack the operator picked up in their second week and has never once said out loud.
What it costs you to decide late
If your robot needs your engineers to teach it every new behavior, you have built a services business with a hardware invoice attached. Every site is a deployment. Every new part number is a ticket. Your margin becomes a function of how many of your people are standing in someone else's factory. That business can work. Walden raised $300 million, which buys a lot of engineers in factories, and having Toyota as both lead investor and first plant makes the early deployments unusually friendly.
If the operator can teach it, you're selling a tool, and your cost to serve drops. You inherit a harder design problem in exchange. The teaching interface has to work for someone who will never read the manual, and a robot that learns from a human learns the bad habits with the same fidelity as the good ones.
Neither answer is free. Both of these teams picked one on purpose, which already puts them ahead of most of the hardware companies I see.
Dave's take
I've watched teams spend a year building what the spec described and about a week finding out what the customer needed. In robotics that mistake hides inside the training interface, because founders file it under AI and hand it to the research team when it belongs to product definition. Decide who owns the behavior after the sale and you've decided what business you're in.
From Dave’s video library
Dave walks through the NBER study of an AI copilot dropped into a 5,179-agent support org, where the newest reps gained the most and the veterans gained almost nothing.
I’m here to help you scale.
Work With DavePrefer a smaller first step? Book a $500 one-hour working session →
Dave Saunders is the founder of Base Reality Group and a Fractional CPO for hard-tech founders. He was a founder and operator at Galen Robotics, where the surgical-robotics platform earned FDA De Novo authorization in 2023, and he managed a 35-patent portfolio licensed from Johns Hopkins. He wrote Founders Who Finish and publishes The Build. More about Dave →