Leadership1 distinct publisher3 min readPublished
Four ex-DeepMind engineers say robotics pilots stall on deployment rather than capability, and their answer moves the retraining work onto the buyer's payroll, which changes what an automation vendor has to prove.
The Board Room · Leadership desk
Compiled by The Board RoomSomething wrong?How this is made
The single quantified claim on the record is a teaching time, and it repays conversion. Reimagine says that at a company extracting critical materials from used hard drives, the time needed to teach a robot a new task fell from one day to 10 minutes [8]. Read a day as an eight-hour shift and that is 480 minutes against 10, a 48-fold reduction [1]. What the figure does not carry is how many show-watch-correct cycles are needed before the arm can be left alone, or what a poor correction costs further down the process. It is the vendor's own number from one site, and it describes an interface rather than an output.
The board-deck version of this company is that customer teaching closes robotics' data gap. The gap Scholz cites comes from Ken Goldberg at UC Berkeley: the largest reported robot-training dataset holds roughly one year of experience, against the 100,000 years a person would need to read and view everything behind leading AI models [6][7], a ratio of about 100,000 to one [3]. Ten minutes of arm-guiding at a recycling plant does not touch that. What it touches is the cost of adapting whatever generalisation the model already has to one site's particular parts and fixtures. The bet is about where adaptation labour sits, not about the corpus.
The tradeoff is straightforward: Reimagine is moving post-training labour off its own payroll and onto the customer's. Scholz's account of DeepMind is that customers called and his team flew out to fix the robots [9], and his named failure mode is the robot that becomes an expensive paperweight when nobody on site can repair its behaviour [10]. Removing the fly-in dependency removes a cost the vendor carried and a delay the buyer suffered. It also books hours against operators who already have a job, and Scholz says that at the recycling site staff went further than that, generating their own use cases for the arms [15].
One less generous reading is that Reimagine is selling an unfinished product and billing the gap as a feature. The evidence pointing the other way is circumstantial: three robotics projects Scholz built at DeepMind that never got past the pilot phase, and a founder who concluded the binding constraint was usability by the staff who understand the work rather than model capability [3][4]. Circumstantial evidence about the deployment layer is still evidence about the layer where automation money gets stranded. What the record does not tell us is whether the behaviours a customer's staff accumulate travel with the customer or stay with the vendor, and that is the term that decides who holds leverage at renewal.
The decade-scale question is separate from the quarter-scale one. Whether robotics gets its own ChatGPT moment, and whether it arrives on two legs from Tesla, Figure or 1X [11][12], is a question about the decade. Whether your next automation pilot survives a shift change is a question about the quarter, and it turns on who can retrain the machine on a Tuesday without raising a purchase order. Scholz expects a downstream economy of robot trainers to grow into that space [13]; if it does, the teaching becomes the priced item in an automation contract, and buyers who signed for hardware alone will find they bought the cheaper half.
Ranked by verification strength, evidence, and original report placement.
Reimagine Robotics was founded last year by Jonathan Scholz, who previously led DeepMind's applied robotics team, with former Google colleagues Oleg Sushkov, Akhil Raju and Misha Denil.
The London- and Sydney-based startup emerged from stealth earlier this month and is backed by VC firms Fly Ventures and Firstminute Capital.
Reimagine has adopted a "monkey see, monkey do" approach for its fleet of robot arms and assemblers, some surface-mounted and some on wheeled platforms: customers teach new behaviour by showing a task, watching the robot attempt it, then correcting it by physically manipulating the arm.
The largest reported robot-training dataset contains roughly one year of experience.
Scholz cited the "100,000-year data gap", a term coined by UC Berkeley roboticist Ken Goldberg, who estimates it would take a person approximately 100,000 years to read and view all the text and images used to train leading AI models.
Scholz said that when he was at DeepMind, customers would have to call and his team would fly out to fix the robots, and that from the customer's perspective it would be better not to have to call.
Distinct publishers with included, body-backed reporting in this cluster.
businessinsider.com
1 article · August 29, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
Chinese carmakers are funding humanoid robots out of the factories they already own1 distinct publisher
leadership
A 2.85x bill of materials: what the FCC's robot ban does to an automation budget1 distinct publisher
leadership
A humanoid costs $46,000 to build, or $131,000 without China. That is the de-risking bill.1 distinct publisher
invest
XPeng's $900m robot spin-out comes with a seven-year clock and a $1bn put2 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One interview, one voice
Two things in this story could be checked by a stranger: who founded the company and who funded it. Everything else — three stalled DeepMind projects, the recycler's 10-minute retraining, workers eagerly inventing tasks — is Scholz describing Scholz's world in his first interview out of stealth, relayed by Business Insider without a customer, an engineer or a second desk to push back. Even the data gap arrives at one remove, as Ken Goldberg's estimate quoted by the founder who finds it useful.
One described site, none named
Commercial traction amounts to a single deployment whose customer is described only by what it recycles, plus 'several' manufacturing tests with no count, no contract, no fleet size and no repeat order. For a company weeks out of stealth that is unremarkable — but it means the deployment thesis is currently supported by one anecdote and a plural adjective.
Modest pitch, one oversized number
Credit where it is due: Business Insider points at the humanoid demo cycle rather than joining it, and Reimagine's own claim is unglamorous — make the thing teachable, not intelligent. The overshoot is narrow and specific. A 48-fold collapse in teaching time is doing persuasive work no verification supports, and a 'downstream economy' of robot trainers is presented as an emerging market when it is one founder's forecast with no timeline attached.
The vendor picked the number
A first post-stealth interview is a recruiting and fundraising instrument, and it shows in what got measured: the one statistic in the piece was selected by the company it flatters, about a customer it declines to name. Two VC backers are named while the cheque is not, which is the disclosure pattern of a firm managing signal. Business Insider's own incentive cuts the same way — an exclusive is worth more vivid than hedged.
Firm on framing, soft on results
We can be fairly sure what Reimagine says it is doing and why — the quotes are direct and the positioning is coherent. We have almost no purchase on whether it works, because every performance claim is single-sourced to the seller and nothing in the story has yet met a second reporter. Expect the founding and funding facts to hold and the teaching-time number to move if anyone checks it.