Robot builders grapple with data scarcity as physical AI reaches "GPT-2 moment"

Physical AI startups are drawing billions in venture funding, yet they face a stubborn obstacle: the field still lacks the high-quality training data and mature tools needed to create robots capable of reliable, commercially viable work. The tension became apparent at last week's Actuate conference in San Francisco, where the robotics AI community gathered to wrestle with how to move beyond the current technological plateau.

The Data Crisis Holding Back Progress#

The challenge facing robot developers today echoes what large language model researchers faced before their breakthroughs. Startups trying to build machines that can manipulate objects or operate in unstructured environments are running into a shortage of diverse, real-world training data. Unlike autonomous vehicles—where companies can collect footage from thousands of human drivers—robotics lacks a comparable data engine.

Harry Mellsop of Antioch, a simulation platform for AI model builders, characterizes physical AI as stuck in its "GPT-2 era"—the pre-ChatGPT phase when more compute and data were needed to unlock a leap forward. The bottleneck is particularly acute for visual and lidar sensor streams, which generate massive datasets. To break through, developers will need not just more data but also better training methods and GPUs optimized for ray tracing to create high-fidelity simulations.

Autonomous Vehicles Lead, Humanoids Lag#

The one robotics domain that has advanced meaningfully is autonomous vehicles. Self-driving teams can tap into real-world deployment data, and the primary task—avoiding collisions—differs fundamentally from manipulation work. This advantage has prompted major car companies to pivot toward humanoid robotics as a natural extension of their AI infrastructure investments.

Tesla's Optimus project is the most prominent example, but now Wayve and Uber have launched robotics labs to explore humanoid form factors. Alex Kendall, CEO of Wayve, argues that the data and machine-learning infrastructure developed for vehicles will carry over to general manipulation tasks—though specific world models may need adjustment for different robot bodies.

The Vertical vs. General Purpose Dilemma#

A sharp disagreement has emerged over strategy. Some builders argue for vertical specialization: Gritt is automating solar farm installation, Agility is deploying robots in warehouses, and Bedrock is operating excavators autonomously. These focused efforts generate real revenue and valuable on-site data. Yet as Théophile Gervet of Genesis AI points out, a robot that succeeds 80% of the time at one narrow task creates no competitive moat if a rival builds on top of a more advanced underlying model.