Here is a fact that sounds like a paradox: the fastest way to teach a machine about reality may be to never show it reality at all. Duality AI has spent since 2018 turning that paradox into a product.
The problem Duality AI is trying to solve is boring in the way that important problems often are. Modern AI - the kind that lets a warehouse robot recognize a box, or a drone-detection system pick an aircraft out of the sky - learns from data. Lots of it. Perfectly labeled, endlessly varied data. And in the physical world, that data is expensive, slow, and sometimes dangerous to collect. You cannot crash a thousand real forklifts to teach a robot what a near-miss looks like. You cannot fly ten thousand real drones past a sensor just to build a training set.
So Duality's bet is this: don't collect the data. Simulate it. Build a digital twin - a virtual replica of the robot, the sensor, the warehouse, the desert, the drone - that is accurate enough in appearance and physics that an AI trained inside it will behave the same way when it steps into the real world. Then generate as much perfectly-labeled synthetic data as you need, at whatever scale you can afford compute for.
FalconCloud is the Google Docs of digital twin simulators.— How Duality describes its browser-based platform
The product that does this is called Falcon, and it is built on Unreal Engine - yes, the same game engine behind blockbuster video games and, increasingly, Hollywood visual effects. This is not an accident. The whole reason to build simulation on a AAA graphics engine is that photorealism, in this context, isn't vanity. It's function. If a synthetic image doesn't look convincing enough to fool the AI, the AI won't recognize the real object when it finally sees one. Graphics quality, in other words, is a safety feature.
The sim2real gap
There's a term of art here that Duality uses constantly: the sim2real gap. It's the difference between how a model performs in simulation and how it performs in reality, and it is the quiet bottleneck of the entire robotics-AI field. Close the gap, and simulation becomes a legitimate substitute for real-world testing. Leave it open, and your beautifully-trained model falls apart the moment it meets a real shadow, a real reflection, a real smudge on a real lens.
Duality's argument is that its Falcon suite delivers AI model accuracy that real-world data alone cannot provide - not just cheaper data, but better-covered data, including the rare edge cases that almost never show up in a real recording but absolutely will show up in deployment.
From Pixar to the Pentagon
The founders make the strategy legible. Apurva Shah, the CEO, came out of Pixar, where he worked on the studio's Oscar-winning film pipelines - the world of making pixels look real. Mike Taylor, the CPO, came out of robotics: he led field-robot teams at Caterpillar and helped win the DARPA Urban Challenge with Carnegie Mellon - the world of making machines move through reality. Put the graphics person and the robotics person in a room, and you get a company whose entire thesis is that those two disciplines were always meant to be the same business.
That pedigree shows up in the customer list. Duality's Falcon platform has been used by DARPA, the U.S. Army, NASA's Jet Propulsion Laboratory, Honeywell, KEF Robotics, and Procter & Gamble - a range that runs from consumer-goods manufacturing to national defense. In 2025, the U.S. Army contracted Duality to take a "digital-first" approach to building an AI-based anti-drone system: instead of collecting endless real footage, simulate the drones and train the detector inside the twin.
It is a small distributed company - roughly 32 people, spread across three continents, holding more than 70 patents across robotics, simulation, and visualization. It raised a $12 million Series A in 2021 after early seed backing. By the standards of the AI-hype cycle, that is a modest war chest. But Duality is not selling a chatbot; it's selling the unglamorous infrastructure that other people's robots are trained on. The most interesting AI companies often work on the plumbing nobody sees.