General-purpose robots will need roughly 10 million hours of high-quality physical interaction data before they can work reliably at production scale, according to a Goldman Sachs report published on 30 October. The estimate, drawn from an expert call with a robotics algorithm engineer the bank held on 28 October, puts a concrete number on a problem the industry has mostly described in slogans: the frontier AI models that can now reason about the world still have almost no record of how to touch it.
The gap between that figure and the money being staked on robots is the story. Morgan Stanley projects as many as 1 billion humanoid robots on the planet by 2050 and a market worth $7.5 trillion. Tesla says its Optimus line could become its best-selling product, with Elon Musk targeting output of 20,000 units a week, roughly a million a year, by the end of 2027. The investment case assumes machines that can already do useful physical work. Goldman's data requirement describes machines that largely cannot, not yet.
What that looks like on the ground was captured in a New York Times report on 27 October. Dimensional, a San Francisco robotics startup building an operating system for physical AI, advertised jobs paying $20 an hour, 39 cents above the city's minimum wage, for people to walk its four-legged robots through the streets. The robots, made by China's Unitree and costing about $2,000 each, cannot yet navigate real pavements reliably, so humans guide them while onboard sensors record the data. Founder Stash Pomichter was blunt about why: AI agents still struggle to control hardware, and "humans are still needed in the loop." A machine meant to replace physical labour currently depends on physical labour to learn.
Goldman's report lays out why the last stretch is so hard. Unlike self-driving cars, which mostly steer, accelerate and brake, a humanoid must manage dozens of mechanical degrees of freedom, continuous contact, shear forces and tactile feedback. The bank flags three bottlenecks: 3D spatial perception that is not yet precise enough for millimetre-level manipulation, inference that is too slow for real-time balance and recovery, and the cost of running capable models on hardware small enough to sit inside a robot. It judges general-purpose manipulation to be about where autonomous driving stood before 2015. First-person video teaches intent but carries no force data; simulations struggle to reproduce real friction and material deformation.
The report is not uniformly bearish, and that is where sources pull in different directions. Goldman expects progress in physical AI to run two to four times faster than autonomous driving did, helped by foundation models that generalise to unfamiliar tasks, and it sees nearer-term money in collaborative and industrial robots rather than humanoids. Academics are pushing the humanoid case regardless: Utah State University's Mario Harper argues that because the built world is shaped for human bodies, only a humanlike machine with reasoning can eventually clear a room or clean a house. Both can be true. The open question the Goldman number sharpens is timing, not direction: whoever assembles 10 million hours of good interaction data first may matter more than whoever ships the most impressive demo.