Physical AI’s bottleneck shifts from what robots can do to whether factories trust them

AI is leaving the data center and running into the messier world of factory floors. That is where the neat promise of physical AI gets tested: not in demos, but in places where conditions change, parts vary and no one wants a robot that freezes the moment reality shifts.

Adrien Gaidon, co-founder and chief strategy officer of Walden Robotics, says the old industrial model is the problem. Traditional robots sit behind cages, repeat one task and were never built for constant change. Manufacturers, he argued, now want general-purpose machines that can switch jobs and keep improving as the process evolves. The hard part is not the idea. It is deployment.

Walden spun out of Toyota Research Institute in January 2026 and launched with $300 million in July. Its robots have already worked full production shifts alongside people since May, doing machine tending, parts kitting and subassembly. Gaidon framed that as a response to work that factories either could not automate or had never tried to automate at all.

CoreWeave is leaning into the same shift from another angle. John Mancuso, the company’s vice president of field engineering, said robots need a lot of simulation before they ever do anything physical, and that infrastructure is part of the offering. But simulation alone does not finish the job. Walden pairs autonomous robots with remote human assistants, and Gaidon says the machines are designed to call for help when they need it.

That human help is not a weakness in his telling. It is the data stream that matters most. Gaidon compared the setup to automated looms from 120 years ago, when one person watched over 40 machines. The point, he said, is that robots can work on their own and still learn from the moments when they need assistance. In manufacturing, trust now seems to depend less on what a robot can do on paper and more on whether it can adapt when the line changes under its feet.