Nvidia Bets on Simulation to Give Robots Their ChatGPT Moment

Since ChatGPT’s public debut in November 2022, generative artificial intelligence has rapidly reshaped software development, content creation, and digital workflows. Physical AI—the integration of machine learning into physical robotic systems—has not yet undergone a comparable inflection point.

One obstacle is the lack of a unified, comprehensive dataset for general-purpose robotics. Large language models can train on trillions of words available across the open internet, while robotics developers face a severe deficit of real-world training data. These technical challenges and the strategies emerging to address them will be discussed at TechCrunch Disrupt 2026 in San Francisco.

Autonomous vehicle operators have traditionally built their datasets through physical-world training. Alphabet Inc.’s Waymo, for example, has accumulated road miles for more than a decade. Collecting an equivalent amount of data for general-purpose humanoid or multi-form robots is slow, expensive, and logistically impractical.

A growing ecosystem of robotics startups is instead developing “embodied AI” through simulation. Engineers construct digital twins of real-world environments with highly accurate physics engines, allowing robots to practice manipulation, navigation, and coordination tasks millions of times in parallel. The resulting synthetic data is intended to train foundation models.

Nvidia Corp. has made itself a central infrastructure provider for this simulation-based approach. Under CEO Jensen Huang, the semiconductor company has expanded into specialized robotics development platforms, including the Nvidia Isaac platform and Omniverse for virtual world simulation.

The method is known as “simulation-to-reality” (Sim-to-Real) transfer. It is intended to replicate the scale of internet-wide text datasets without requiring physical machines to learn every task through trial and error.

Nvidia’s portfolio also includes Project GR00T, a foundation model designed for humanoid robots to understand natural language and emulate human movements. Les Karpas, the Global Head of Physical AI for Nvidia Inception, coordinates the company’s relationships with startups building on these platforms.

Nvidia Inception operates as a virtual accelerator, providing technological support and venture connections to early-stage AI and hardware firms. Karpas’s background includes architecture, manufacturing engineering, and corporate venture capital roles at iRobot, Stanley Black & Decker, Herman Miller, Intellectual Ventures, and Cirque du Soleil.

At TechCrunch Disrupt 2026, scheduled for October 13–15 at San Francisco’s Moscone West, Karpas will speak on the Real World AI Stage in a session titled “Robots Are Waiting for Their ChatGPT Moment. Here Is What Is Standing in the Way.”

The session will include founders and executives from several parts of the physical AI landscape. Shield AI is developing autonomous AI piloting systems through its “Hivemind” software, which enables military aircraft to operate autonomously in GPS-denied environments without human pilots.

Foxglove develops visualization and debugging tools for robotics engineers. Its products help teams manage and interpret complex sensor and diagnostic data during development, while FieldAI focuses on autonomous capabilities for mobile robots working in unstructured, rugged, and unpredictable physical environments.

Colossal Biosciences uses advanced hardware automation and laboratory robotics in its genetic engineering and de-extinction research workflows.

The three-day conference is expected to attract more than 10,000 technology founders, investors, and industry leaders. Alongside its physical AI sessions, the event is a venue for startup fundraising and corporate development.

Registration pricing is scheduled to increase after September 25, while organizers are offering group discounts of up to 30% for corporate and startup delegations.