Before a robot meets a dirt road, a warehouse aisle, or a camera pointed into bad weather, Apurva Shah would like it to spend time in a world that can be reset. In that world, a team can move the sun, change the terrain, introduce an unlikely obstacle, and run the same encounter again. The robot can fail without bending metal. Its developers can learn without waiting for reality to produce the precise mistake they need to study.
This is the practical promise behind Duality AI, the San Mateo company Shah co-founded with roboticist Michael Taylor in 2018. Its Falcon platform builds digital twins of machines, sensors, and environments, then uses those virtual systems to generate synthetic data and test autonomous behavior. The company describes its purpose in five economical words: “Virtual worlds for solving real problems.”
The phrase also summarizes Shah's career. Long before “physical AI” became an industry label, he was learning how to make complicated digital worlds hold together. His route went through electrical engineering, computer science, computer graphics, animation, product design, teaching, and startup building. Each stop changed the audience for the world on the screen. The work itself kept returning to a familiar question: can art, code, and systems cooperate well enough to make a constructed environment useful?
The pixels had a job
Shah began his professional career in 1993 at Ringling College of Art and Design, working as an instructor and systems administrator. A year later he joined PDI, which became part of DreamWorks, and worked on the pipelines behind early computer-animated films. His credits from that period include Antz and Shrek. In 2001, he moved to Pixar.
At Pixar, Shah's work ranged across effects, technical development, and production leadership. He contributed to Finding Nemo, served as effects supervisor on Ratatouille, worked in global technology on Toy Story 3, and was supervising technical director on Cars 2. He also co-authored research on turning unruly particle simulations into coherent surfaces, the sort of behind-the-scenes problem that lets water, food, smoke, and other difficult material behave consistently from frame to frame.
One small Cars 2 story reveals his instincts. Director John Lasseter needed to review work from lighting, animation, effects, set design, and layout while away from the studio. Shah wrote an iPad application that let Lasseter record voice-memo feedback for those departments, including during his commute. It was not a shot in the movie. It was a tool that made the system around the movie move more smoothly.
“Context is increasingly playing a bigger and bigger part in products.”Apurva Shah, 2017
That preference for tools over spectacle helps explain what came next. After Pixar, Shah held global technology leadership roles at Prana Studios and Rhythm & Hues. He founded Whamix, a creative studio and technology lab, and later led creative technology at Capital One, where his teams worked with immersive media, artificial intelligence, connected devices, and 3D game engines through a human-centered design lens.
In a 2017 interview, he said visualization was close to “breaking out of the screen.” Graphics, in his telling, would no longer be confined to a desktop or phone. They would become immersive and mingle with the physical world. It sounds prescient now, but it was also continuous with his work: a belief that images become more consequential as they acquire context and invite action.
Two founders, two halves of a world
Duality paired Shah's background in graphics and production systems with Taylor's experience in field robotics. Taylor had led Caterpillar teams building large machines and had worked with Carnegie Mellon's winning entry in the 2007 DARPA Urban Challenge. Shah understood how to construct high-fidelity environments. Taylor understood what physical machines demand from them. Their overlap was simulation.
The initial challenge was specific: accurately simulate complex environmental scenarios and observe how sensors and machines respond. The task sounds like animation, except the standard is different. A scene can look convincing to a person and still teach a machine the wrong lesson. A camera model needs the right optics. A vehicle needs the right dynamics. Terrain needs more than texture. The data needs labels, provenance, and controlled variation.
Falcon grew around those requirements. It uses Unreal Engine for real-time environments and supports digital twins that encapsulate appearance, physics, behavior, and sensor response. Shah has offered a web metaphor for the stack: Unreal Engine is the browser, Universal Scene Description is HTML, and Python is JavaScript. The comparison turns a specialist's simulation environment into something modular and legible - a world assembled from structured parts rather than fused into one opaque file.
Physics-based simulation
Precise geometry, behavior, sensors, and controlled initial conditions.
Generative world models
Broad variation and accessible creation, with less direct control over grounding.
A laboratory for the unlikely
Simulation matters most when reality is stingy. A team developing a drone detector cannot summon every aircraft, weather pattern, viewing angle, and background on demand. An off-road autonomy group cannot cheaply replay the exact combination of slope, soil, vegetation, and sensor placement that caused a failure. A factory cannot keep interrupting production to create new defect images. Digital twins let teams specify those conditions, generate data, and repeat an experiment.
That logic has put Duality into projects involving NASA's Jet Propulsion Laboratory, DARPA programs, the U.S. Army, Honeywell, Autodesk Research, and Kitware. The applications vary: off-road navigation, manufacturing inspection, drone detection, robotic assembly, and frameworks for evaluating autonomous-system behavior. The connecting tissue is the ability to examine a machine inside a controlled context before trusting it outside one.
Shah's public writing stays close to that applied frame. In 2023, he argued that current language and multimodal models were already capable enough to tackle narrow autonomy problems when combined with predictable parts of the traditional robotics stack. In 2025, as generative video and world models accelerated, he drew a distinction between implicit models, which offer variety and easy creation, and explicit simulators, which directly encode geometry and physics.
His answer was not to choose a camp. He proposed hybrid pipelines. Generative tools can broaden and accelerate world creation; physics-based simulation can ground those worlds and provide control. The ambition is to close what he calls the Gen2Real gap - the distance between data that looks plausible and data that causes a physical system to behave correctly.
“Hybrid synthetic data pipelines and agentic workflows can combine the strengths of simulated and generative approaches.”Apurva Shah, 2025
The teacher inside the engineer
Alongside the company work, Shah taught interaction design at California College of the Arts. In 2021, before the metaverse became a blanket term for almost every virtual experience, he taught a course called Designing the Metaverse. He defined the concept through interaction: a shared, persistent three-dimensional context in which embodied users navigate and act through direct presence.
The definition matters less as a prediction than as a method. Strip away the brand names, then identify the system's durable parts. Who acts? What do they perceive? What persists? Which feedback loop connects action and environment? The same habit appears in his robotics work. A digital twin is useful because the scene can be decomposed, inspected, changed, and understood.
Even Shah's lighter public observations return to this theme. Writing about Sunday hikes, he described being awed by the variety and complexity of outdoor biomes, then turned to the challenge of building virtual versions suitable for cameras, infrared, thermal imaging, radar, and synthetic-aperture radar. The walk becomes a systems diagram. The landscape is beautiful, but it is also radiometry, geometry, material, weather, and scale.
From believable to dependable
In February 2026, Duality introduced Vibe Sim, an agentic copilot for creating simulations through prompts. A month later, Shah co-authored an article on virtual synthetic-aperture-radar sensors and the digital worlds needed to make their output useful. In June, he presented world modeling and vibe simulation at Unreal Fest as Duality joined Epic Games' Unreal Engine Service Partner Program.
These updates place his oldest craft inside the current AI cycle. The interface is becoming conversational. The worlds are becoming easier to assemble. The obligation to ground them remains. When an image is entertainment, a small physical mistake may pass unnoticed. When the image becomes training data for a machine, the mistake can propagate.
Shah's career makes that shift unusually visible. At Pixar, pixels served a story. At Capital One, emerging technologies served a customer experience. In a classroom, virtual-world frameworks served collaboration. At Duality, simulation serves an experiment, and the experiment serves a machine that will eventually leave the screen.
The most interesting part is not that a visual-effects engineer moved into robotics. It is that the same discipline kept acquiring a stricter definition of success. First the world had to be believable. Then it had to be interactive. Now it has to teach, test, and transfer. The robot gets to fail first. The virtual world has done its job when the physical one is less surprising.