Physicl is Helping Digitize Real World Objects for Simulation

“Digital cousins” is a new one to me. A quick Google search surfaces a research paper from 2024, coauthored by AI pioneer, Fei-Fei Li, titled “Automated Creation of Digital Cousins for Robust Policy Learning.” The research addresses the expense and other issues involved in generating digital twins, proposing as a stop-gap, “the concept of digital cousins, a virtual asset or scene that, unlike a digital twin, does not explicitly model a real-world counterpart but still exhibits similar geometric and semantic affordances.”

So, digital assets that are, you know, related, but not that related. That is to say that, when constructing simulation, there are those times when pretty close has to be close enough. And hey, if it was good enough for Patty Duke, who’s to say it’s not good enough for the occasional world model?

From what I can gather, two and a half years later, the term has yet to catch the world ablaze, though Alex de Vigan did casually drop it in conversation last week, which is precisely why we spent the last two paragraphs talking about the thing. He was, of course, referring to its use – along with digital twins – in the context of his recently announced startup, Physicl. 

“Basically, you have inputs in a real object, but you don't have all the inputs to create the exact specific,” he notes, contrasting the digital relatives. “So there is going to be from 10-20% approximation on some of the inputs. Based on that level, depending on the type of input, you can create the different types of output.”

This specific aspect of Physicl’s work, in turn, came up during a conversation about the company’s fascinating history. The startup is a kind of spinoff of existing company, Nfinite – a kind of company cousin, if you will. It was founded by members of that team, including de Vigan, who remains CEO of both. 

Nfinite was built around utilizing 3D data to build spatial models. “From enhancing e-commerce with stunning product imagery to supporting leading tech firms and foundational models with IP-free, metadata-rich spatial data,” the company writes, “we empower our customers to advance visual AI to the next level.”

In January, Nfinite announced  a high-profile collaboration with Getty, aimed at transforming the wire service’s massive image catalog into 3D assets. 

“The partnership marks a significant step forward in bridging the gap between traditional visual media and the data infrastructure required to build and train spatially-aware, multi-modal AI models able to perceive, understand, and interact with the physical world autonomously,” the companies noted at the time.

de Vigan, who had previously worked as an M&A lawyer, says he was missing one key piece when he founded Nfinite back in early 2017. 

“Ten years ago, my vision was to be able to digitize the world in 3D,” he says. “I had this very big intuition that it would be transformative for a lot of industries. Thought about creative, thought about commerce, thought about education. Unfortunately, did not think about robotics. But 10 years later, we built these capabilities, we've applied it successfully to retail, and now we decided to focus on physical AI.”

As it emerged out of stealth, Physicl had no robotics customers to speak of, but the company was hoping that a high profile launch at GTC would raise interest in the company’s growing catalog of digital assets. 

“What we are is a data layer, beneath all the different types of data that you might have – images, video, blueprint, point cloud, 3D. Our job is to normalize that into super high fidelity, standardized physical data that is actionable in your training frameworks. We do not do the simulation engine. We sit on top of Isaac Sim, for example.”