Robots Are Waiting for Their ChatGPT Moment: What Stands in the Way

General-purpose robots still lack the internet-scale datasets that enabled large language models to move from research to everyday use after ChatGPT’s 2022 launch. Nvidia Inception’s Global Head of Physical AI, Les Karpas, will address this gap and potential solutions during his session on the Real World AI Stage at TechCrunch Disrupt 2026.

Why general-purpose robots still lack a defining breakthrough moment

The release of ChatGPT in November 2022 demonstrated how access to vast language datasets could turn research models into widely used tools. Robotics has accumulated decades of incremental progress yet has not reached an equivalent inflection point. According to the session description, the missing element is an internet-wide dataset for physical interactions comparable to the text corpora used by OpenAI and Anthropic. Self-driving programs such as Waymo accumulated data through years of real-world miles, but this approach remains limited to narrow domains and geographies. Karpas identifies the absence of broad, diverse physical data as the core constraint preventing general-purpose robots from achieving similar capabilities. The world has not seen robotics achieve the same leap because no comparable corpus of everyday physical interactions exists at internet scale. Even accumulated road miles from autonomous-vehicle fleets cover only specific environments and continue expanding market by market rather than delivering universal coverage. Without that breadth, general-purpose robots cannot train on the variety of scenarios needed to operate reliably outside controlled settings. The all-too-real problem facing robotics is that nobody has an internet-wide dataset for physical AI in the way language models had for text.

How simulation and synthetic data are being used to close the gap

Without a ready-made physical dataset, companies are attempting to create scale artificially. Approaches include running large numbers of simulated environments, generating synthetic interaction data, and training foundation models across multiple robot morphologies simultaneously. These techniques aim to replicate the breadth that language models obtained from web text. The session will explore whether these methods can produce the robustness required for unstructured real-world settings. Karpas will outline both the promise and the current shortcomings of these strategies based on his work with the robotics startup ecosystem. A growing number of startups now focus on manufacturing equivalent scale through simulation, synthetic data, and foundation models trained across many robot forms at once. The challenge lies in bridging digital and physical worlds so that models trained in simulation transfer safely to unstructured environments. Karpas will examine how far these artificial-scale methods have progressed and where gaps remain before they match the robustness language models gained from naturally occurring web data. The same companies trying to solve that data gap are the ones he regularly interfaces with through Nvidia Inception.

Nvidia’s view of the physical AI opportunity and constraints

Nvidia has repeatedly highlighted robotics as a major growth area in recent keynotes. Karpas coordinates the company’s relationships with startups operating in robotics, automotive, manufacturing, mobility, and smart cities. His role gives him direct visibility into the data-generation efforts underway across these sectors. The company’s perspective emphasizes that solving the physical data bottleneck is a prerequisite for scaling capable general-purpose systems. Karpas will discuss how Nvidia sees the interplay between hardware acceleration, simulation platforms, and the startup innovations needed to generate usable datasets at sufficient volume and variety. Nvidia’s bullish stance on robotics stems from the belief that once the data gap narrows, new markets will open rapidly. Karpas interfaces regularly with the same startups attempting to close that gap, giving him a current view of both technical progress and remaining commercial hurdles. Anyone who has seen Nvidia CEO Jensen Huang’s keynotes the past few years knows the company is bullish about the robotics field.

Les Karpas’ background and relevance to the robotics data problem

Karpas’ career spans architecture, manufacturing engineering, startup leadership, venture studio roles, and corporate venture capital. Previous positions include work at Stanley Black & Decker, Intellectual Ventures, Herman Miller, iRobot, and Cirque du Soleil. This range of experience across physical product development, robotics platforms, and investment provides context for evaluating both the engineering and commercial dimensions of the dataset challenge. His session will draw on these perspectives to assess what types of data and collaboration models are most likely to advance the field. The cross-disciplinary path equips him to evaluate not only technical feasibility but also the investment and partnership structures required to scale physical AI datasets. His path to getting there is one that speaks to the cross-disciplinary skills the robotics industry demands.

Real World AI Stage sessions at TechCrunch Disrupt 2026

The Real World AI Stage was created to examine the specific difficulties of bridging digital models with physical systems. In addition to Karpas, founders from Shield AI, Colossal Biosciences, FieldAI, and Foxglove will participate in related discussions. These sessions focus on the practical obstacles and emerging opportunities in applying AI to real-world environments. The stage programming reflects the recognition that physical AI requires distinct data strategies and validation methods compared with language or vision models trained on internet content. Attendees will hear how multiple organizations confront the same dataset limitations from different industry angles. Karpas will be joined by these founders as they walk through the unique challenges and opportunities they face.

Frequently asked questions

What is the main limitation preventing general-purpose robots from advancing rapidly?

The primary constraint is the lack of an internet-scale dataset capturing diverse physical interactions. Language models benefited from vast text corpora; equivalent data for robots does not yet exist at comparable breadth or volume.

How are companies attempting to create large physical AI datasets?

Startups are using simulation environments, synthetic data generation, and foundation models trained across multiple robot forms to manufacture the necessary scale artificially.

Why does Nvidia consider robotics a major focus area?

Nvidia executives have stated that advances in physical AI could open substantial new markets once the data bottleneck is addressed, and the company maintains active partnerships with robotics startups through its Inception program.

What experience does Les Karpas bring to this discussion?

Karpas has held roles in manufacturing engineering, robotics product development, startup leadership, and venture investment, giving him visibility across both the technical and commercial aspects of physical AI.

When and where will the session take place?

The session is scheduled for TechCrunch Disrupt 2026, held October 13-15 at Moscone West in San Francisco.

Key takeaways

• Robotics lacks an internet-wide physical interaction dataset equivalent to the text data used for language models.

• Simulation, synthetic data, and multi-robot foundation models are the main approaches being pursued to build scale.

• Les Karpas coordinates Nvidia’s relationships with startups working on these data challenges across multiple industries.

• Karpas’ background spans manufacturing, robotics companies, and venture roles relevant to evaluating physical AI progress.

• The Real World AI Stage at Disrupt 2026 will feature additional founders addressing related real-world AI constraints.

Outlook for physical AI progress

The central open question remains whether current simulation and synthetic data methods can deliver the diversity and robustness needed for general-purpose robots. Karpas’ session will provide an update on the state of these efforts from Nvidia’s vantage point within the startup ecosystem. Attendees will hear directly about both the technical hurdles and the collaboration models being tested to accelerate dataset creation. The discussion is expected to clarify near-term milestones and remaining gaps before physical AI reaches wider deployment. This is your chance to hear firsthand how Nvidia is thinking about the single biggest bottleneck in physical AI, and what it’ll take to clear it.