On a TED stage in 2019, Ivan Poupyrev moved a slide by brushing the sleeve of his jacket. The gesture was tiny, almost private. No glowing panel, no controller, no theatrical wave. Conductive yarn woven into the cuff registered his fingers, and the room’s largest screen quietly obeyed. Poupyrev had spent years helping create Jacquard at Google, but the trick carried an older ambition. He wanted computing to stop demanding that people visit a rectangle before anything interesting could happen.
The jacket joined an improbable collection of interfaces from his career: a virtual arm that stretches across a simulated room, a screen that changes how a fingertip feels, a plant that sings when touched, puffs of air that can be felt without contact, and a radar chip that notices movements too small for a camera to need. Some became products. Some remained experiments. Together they form a consistent argument. Digital systems should enter the materials and spaces of ordinary life, while asking less of the people using them.
A question with a long shelf life
Poupyrev grew up in the Soviet Union. After its collapse, scientific funding became harder to find, and in 1994 he left. His studies took him between Japan and the United States. At Hiroshima University, he completed a doctorate on 3D user interfaces for virtual reality, with research done in collaboration with the Human Interface Technology Laboratory at the University of Washington.
One early result was the Go-Go interaction technique, co-authored in 1996. A user could reach naturally for nearby virtual objects, then extend farther as the software lengthened the virtual arm nonlinearly. It was a clever answer to a basic mismatch: a digital room can be enormous, but a human reach is not. Even then, Poupyrev’s subject was not virtual reality for its own sake. It was the seam where a body meets a computational world.
He spent roughly nine years in Tokyo at Sony Computer Science Laboratories. There he worked on tactile interfaces such as TouchEngine and explored how consumer devices might respond through the sense of touch. The laboratory gave him room to invent, but it also exposed the distance between a persuasive prototype and an object that survives factories, supply chains, and customers.
At Walt Disney Imagineering Research, where he later directed a small interaction technology group, the materials became more playful. Botanicus Interacticus turned an ordinary plant into a touch-sensitive interface. A person could pinch a stem or slide along a leaf and produce graphics or sound. REVEL altered the tactile qualities of physical surfaces through electrical signals. Aireal created haptic sensations in open air. The experiments were whimsical on the surface and rigorous underneath. They asked what an interface could be if glass were no longer the default material.
When the prototype must ship
Poupyrev joined Google’s Advanced Technology and Projects group in January 2014. ATAP operated less like a conventional research lab and more like a studio with a clock running: projects were expected to become products or end. For an inventor accustomed to the long afterlife of prototypes, that pressure changed the work.
Jacquard began with a visual analogy. A touchscreen is a grid of electrodes. Textile is a grid of threads. Replace some ordinary yarn with conductive yarn, and touch sensing might become part of the material instead of a device attached afterward. Poupyrev entered ATAP knowing little about fashion, then assembled the engineering and design work needed to manufacture interactive fabric at scale. Collaborations followed with Levi’s, Saint Laurent, Adidas, and Samsonite. A Levi’s commuter jacket eventually entered Cooper Hewitt’s permanent collection.
Soli attacked a different boundary. Its miniature radar captured subtle motion without requiring touch. Machine-learning systems interpreted the dense signals created by a hand moving near a device. The technology reached Pixel and Nest products. Across Jacquard and Soli, the teams’ work shipped in more than 15 products across 33 countries. Poupyrev received Cooper Hewitt’s National Design Award for Interaction Design in 2019.
The product record mattered, but the radar signals created another insight. Sensors do not arrive with explanations. They produce changing measurements that specialists decode, often with a custom model for each machine, task, or environment. Around the same time, large language models were demonstrating that one large model could generalize across many linguistic tasks. Poupyrev and his colleagues began asking whether a related approach could work for the much stranger grammar of the physical world.
The data lake has no vocabulary
In 2023, Poupyrev left Google with four ATAP colleagues: Brandon Barbello, Leonardo Giusti, Jaime Lien, and Nicholas Gillian. They founded Archetype AI around a blunt premise. The physical world is already speaking through sensors, but most of what it says is stored rather than understood.
A factory machine produces vibration, temperature, current, pressure, and sound. A street produces video, traffic counts, timing, and motion. A vehicle combines cameras with radar, location, and mechanical telemetry. The data is abundant, yet meaning often arrives through a dashboard configured for one question or an expert inspecting one stream after the event.
Connect measurements to context, turn context into an explanation, and move the explanation close enough to the asset that a person can act in time.
Archetype’s foundation model is called Newton. It is designed to fuse different types of time-series and sensor data, find patterns, predict physical behavior, and connect those findings to natural language. The company’s 2024 research described a model learning across mechanics, thermodynamics, and electromagnetics from raw measurements without being given the underlying equations. The name fits the ambition, though the practical pitch is less about rediscovering physics than making industrial signals legible.
Newton has moved from a research model into a platform for building what Archetype calls Physical Agents. These applications can monitor a process, verify a task, detect an anomaly, or guide an operator while combining several live inputs. Early work has included construction management with Kajima, traffic monitoring with the City of Bellevue, and edge deployments with NTT DATA and other partners. In November 2025, Archetype announced a $35 million Series A, bringing its disclosed seed and Series A funding to $48 million, and opened its platform and Agent Toolkit to select customers.
Bigger than a robot
Poupyrev is careful to separate his idea of physical AI from the current fascination with humanoid robots. A robot is one form. His larger opportunity is the installed world: turbines, warehouses, electrical systems, roads, appliances, and vehicles that already contain sensors and software. Give those assets a shared way to interpret their surroundings, he argues, and intelligence can be distributed without waiting for a new mechanical body.
That view also changes where the model should run. A delayed cloud response can be useless beside a fast machine, and continuous video or industrial telemetry may be too sensitive or expensive to send elsewhere. Archetype has demonstrated Newton operating on a single local GPU. Poupyrev frames privacy as data sovereignty: the people or organizations generating physical data should control where it travels and how it is used.
His preferred shorthand is “augmentation > automation.” It echoes the human-centered language he has used throughout his career. The point of an interface is not to advertise the machine’s intelligence. It is to extend perception or action without forcing a person to translate every signal by hand.
The same seam, one scale larger
In 2026, Poupyrev returned to Tokyo for a summit and posted an old photograph of himself holding TouchEngine at Sony CSL. The device belonged to another era of computing, when adding convincing tactile feedback to a screen felt like a frontier. He drew a line from that work through radar and sensing at Google to Newton at Archetype. The throughline, he wrote, was that digital intelligence remains incomplete until it can perceive, reason, and act in physical reality.
There is a revealing modesty in the old picture: a researcher, a box, a cable, a prototype meant to make one sense richer. Archetype’s language now spans general models and physical intelligence, but its founder still gravitates toward concrete signals and ordinary objects. A jacket sleeve. A machine vibration. A temperature curve. A person entering an intersection.
Poupyrev’s career suggests that the interface after the screen may not look like a single replacement. It may be a growing ability to ask the physical world a question in familiar language and receive an answer grounded in what its sensors can observe. The work began by stretching a virtual hand. It now aims to stretch human perception across the machinery and environments we already inhabit.