Industrial Next, a start-up founded by two former Tesla engineers, is building self-contained robot stations that learn a task in a simulation and can be moved to a new product in minutes instead of weeks. BMW's technology office is evaluating the stations. The company claims they already do delicate phone assembly for a manufacturer it will not name, so for now the performance is its own claim and not something a customer has confirmed.
Robots fail at change, not at speed
A normal factory robot is programmed for one task on one product. If the part changes a little, the robot can fail, and changing a line takes weeks or months of reprogramming. So manufacturers have learned to design products around what robots can repeat, and to give the awkward jobs to people.
Apple's phone supply chain shows the cost of that rigidity. Foxconn has hired hundreds of thousands of seasonal workers to build each new iPhone, because people can be moved to the next product and a line built for one phone cannot. A robot line sized for the launch peak would sit mostly idle for the rest of the year, and it would need rebuilding when the next model arrived.
A robot can practice before it sees the part
Robots have no huge library of examples to learn from, the way chatbots learned from the internet. Industrial Next's answer is to let the robot practice in a detailed virtual world that copies real physics. It works the way a pilot trains in a flight simulator: by the time the pilot is in a real cockpit, the process is already familiar.
The robot learns the process in general, not one part's exact shape, so it can handle a part it has barely seen. The start-up describes a station that peels protective film off camera modules and buttons at human speed or faster, and kept that performance when moved to a phone model it had never seen. New tasks are taught through simulation and demonstration, so the person doing the teaching does not need to write code.
What changes when switching is cheap
Take a maker of electronics that sells a product for about 40 weeks before redesigning it. If the robot cell needs four weeks of reprogramming for every new version, that is 10% of the product's life spent changing over, plus the engineers' time. With a changeover that takes an afternoon, the same cell can serve many short runs and many variants. That is the arithmetic that has kept many small and mid-sized manufacturers from automating at all.
The first jobs to change are therefore the fiddly hand tasks on products that do not last: threading cables, pressing in connectors, peeling film. People kept those jobs because they could adapt, and the motions themselves were never the hard part. The work of reprogramming also changes hands, from an outside integrator to someone on the floor who demonstrates the new task, and product designers no longer have to simplify a design to suit the robot.
The start-up is also not alone in aiming at this. In November 2025, Foxconn and Alphabet's robotics software company Intrinsic announced a joint venture to move from automation built for one product to general-purpose robots that need less re-engineering with each product generation. When the biggest assembler of phones and a start-up both aim at the same weakness, that weakness is clearly what has been holding automation back.
Speed has never decided whether a robot cell pays for itself. The time it takes to change it has. Anyone who is pitched a robot for their line should ask how long a switch to a new part takes, and who has to do it.