Live Recording: the Missing Data for Physical AI

Recorded live at Lumafield’s San Francisco headquarters, this special episode of Go/No-Go features a panel of three founders attacking different bottlenecks in building physical things, hosted by Haomiao Huang, founding partner at Matter Venture Partners. Eduardo Torrealba is the co-founder and CEO of Lumafield, which combines X-ray CT scans and AI analysis technology to give manufacturers unprecedented insight into the products they make. Vikram Pavate’s Tacta Systems trains dexterous robotic hands on data from sensor-equipped work gloves, and Pari Singh’s Flow Engineering builds AI agents that track how a change in CAD, firmware, or simulation can break a requirement elsewhere in a design.

They all agree that the bottleneck in physical AI is data that’s not scrapable from the open Internet, from the motion a worker’s hand makes as it seats an electrical connector to the engineering requirements that go into a great product, which can run into the millions. Each also shares a moment the physical world punched back: an early Lumafield scanner beaten down in an Arkansas factory with no air conditioning, a glove a factory worker rejected after half an hour as too soft, and a decision to throw out every line of code eight years in and start over.

The panel closes on reshoring, the “missing middle” of products that cost $5 to $500 to make, and what the next generation of hardware companies will be able to build.