The expensive part of a factory robot is rarely the arm. It is the weeks you lose when the product changes and the cell has to be taught all over again.
Industrial Next, a San Francisco startup, is pitching a shorter version of that wait. Co-founders Lukas Pankau and Allen Pan, both out of Tesla, say a station can take a new task in minutes rather than weeks or months, without the usual programming cycle. TechCrunch reported the pitch on October 13, 2026, from the company’s Startup Battlefield appearance at Disrupt in San Francisco.
Speed, accuracy, and the phone work are still the company’s account. No independent test or published sample has been released.
“Minutes” is the first thing to unpack. A model update, a fixture change, a physical station move, and the time to get back to rate are different purchases. Do not sign for one because someone described the other.
The complaint behind the company is older than the company. Pankau, Industrial Next’s CEO, was a lead electrical architect at Tesla and later moved to Waymo, according to TechCrunch. Pan led Tesla’s autonomous factory efforts. The company says the two co-founded Industrial Next in November 2021. Pankau’s account of the factory floor, as quoted by TechCrunch:
We had all these fancy things that we could do on cars. I could run machine learning algorithms to design a car for me. But I couldn’t figure out how to program a robot, and this was pretty common for engineers at Tesla. You had this massive divide, right? You had people who are really good at what they’re doing, designing cars and designing self-driving systems, and then you had them effectively stumped at the factory looking at this equipment from the 1980s.
Most factory robots are certain, which is the problem
A conventional industrial robot is programmed for one product and one motion. It is excellent at that motion. Change a clip, a cable, or a model year and the cell can fail, or sit for reprogramming, new fixtures, or a machine swap. Those changeovers often run for weeks or months. The line is not earning while they happen.
Plants do not all answer the same way. Many keep people on the variable steps, or add fixtures and flexible cells. Others simplify the product until a machine can repeat it. Fixed automation stays reliable. In those cases the design gets narrower so the cell does not have to cope. Industrial Next’s argument is that this bargain is outdated. AI automation, they say, should absorb more of the variation instead of forcing every exception onto a person or a fixture.
There is a fair counterpoint, and it is why a lot of that older equipment is still on the floor. A scripted robot that knows one motion is dull. It is also easy to inspect and sign off. A model that is supposed to generalize is flexible, and its misses are harder to write into a control plan. Flexibility is not free just because the demo looks fluid.
The unit on offer is a self-contained robotic station, not an unattended plant in a crate. In operator language, the foundation model they describe is one model tying vision, task context, and motion. A new job is taught by demonstration and simulation, not a hard-coded script for that station alone. Stations, they say, can be reconfigured, combined, and moved across lines. The slide is not the test. What counts is a real changeover: model update, fixture change, station move, and recovery to rate.
The data gap is the serious half. Language models had the open web. A factory has logs, work instructions, and years of cycle history. It does not have internet-scale labeled examples of contact: a cap that almost clicks, film that stretches, a flex cable that fights the gripper. Pankau told TechCrunch, “We can make robots do all these cool things like dance around, because there’s data for that, but we can’t make them assemble high-precision things.”
Industrial Next’s answer is simulation. Pankau said the robots are given a general understanding of the process, built with “an extremely high-fidelity simulation engine that is close to the real dynamics, ” close enough, in his telling, to do high-precision work “almost without seeing the parts at all.” Read that as a claim about how much precision the simulation carries. It is not an argument for switching the cameras off.
This is where robotics demos usually fail. Simulation is clean. Contact is not. Friction, stretchy film, a connector that almost seats: virtual physics is a poor witness. Moving a skill from that engine onto a real part is the whole bet, and nobody outside the company has published how often the transfer misses.
A phone is not a car, and an evaluation is not a plant
What they point to today is handset assembly, not a vehicle line. Pankau has said stations are already used on one of the world’s most popular smartphones. He has not named the customer. The jobs are the fiddly ones that have stayed with people: threading a delicate flex cable, snapping a tiny connector cap, and peeling protective film off camera modules and buttons. On the cable and the cap, he said manufacturers had tried for years to automate the work and failed.
On film peeling, TechCrunch reported him claiming performance at or above human speed, with 100% accuracy, including on a phone model and a production line the system had never seen. AI Weekly’s summary of that reporting is narrower: human speed, across handset models. Neither account includes a sample size, a definition of a defect, or an outside count.
The other item from the stage was BMW. The company said BMW Group Technology Office will evaluate its dexterous manipulation for manufacturing: fine part handling, not a repeated pick. TechCrunch reported no financial terms, no facility, and no timeline.
A rewrite at AI Weekly headlined the same news “Industrial Next Lands BMW Pilot for Foundation-Model Factory Robots That Retool Lines in Minutes.” Pilot is not what TechCrunch described. For anyone weighing AI for business, an evaluation with a serious automaker is a signal. It is not a reference install, and it is not a reason to rewrite a capital plan. No funding figure, headcount, or price has been reported either.
The destination is stated anyway. Pankau told TechCrunch they started the company for the performance they had wanted at Tesla, where stations were bypassed because traditional programming could not keep up. His line for the end state: “We ultimately want to take this to building a full car end-to-end without any human touching ever.”
Set that next to a flex cable. Seating a handset connector is a hard manipulation problem. Building a car is a safety case, a maintenance schedule, a quality system, and a long list of exceptions that still need an owner when the model is wrong. Assembly with no human touch is an ambition. It is not a result, and it sits a long way from the smartphone work the current story rests on.
Questions worth asking before anyone moves a line
- Can these stations really switch tasks in minutes?
That is Industrial Next’s claim, not a published benchmark. Nothing public says whether the clock covers a model update, a fixture change, a station move, and recovery to rate, or only a software swap on a cell already running.
- Did BMW deploy this?
No. BMW Group Technology Office is evaluating dexterous manipulation for manufacturing. No plant, price, contract value, or timeline has been reported.
- Is the 100% accuracy number real?
It is Pankau’s claim, as reported by TechCrunch, not an independent test. There is no published sample size or defect definition behind the percentage.
- What is actually in production?
Unnamed smartphone assembly, as the founders describe it: flex cables, connector caps, and film peeling. A full car with no human touch is the goal, not the evidence.
- Why simulation instead of factory data?
Plants already have logs. They lack internet-scale labeled data on contact, friction, and soft parts. High-fidelity simulation plus demonstration is the company’s answer, and contact is where that transfer usually fails.
- What would “no human touching ever” leave unsolved?
Exceptions, maintenance, quality escapes, and liability. Someone still owns the moment the film tears or the connector does not seat.
If the station idea holds, the payoff is operational. Shorter changeovers. Less pressure to simplify a product so an older cell can repeat it. A path into work that stayed manual because it was too delicate for fixed automation. If it does not hold, the purchase is a convincing practice room with a foundation-model label on the crate.
Treat it the way a careful operator treats any changeover promise. Ask for the line, the sample, the failure, and the person who still walks over when the model is wrong. The complaint about 1980s equipment is fair. Proof that the replacement is ready for a vehicle plant is not in yet. BMW has agreed to look. That is the honest size of the news.