Robotics Training & Data Scarcity

The single biggest obstacle to progress in robotics today is data scarcity.

While software-based AI models have achieved unprecedented scale, robotics — the domain of physical intelligence — is still data-poor.

As shown above, the gap between robotics and other AI domains is staggering. Large language models like GPT-4 were trained on nearly all human text ever digitized — trillions of words. Vision models like Midjourney and Stable Diffusion learned from billions of labeled images and videos.

Robotics, however, has access to only a few million real-world interactions — a dataset smaller than what a single human experiences in their first year of life.

The Data Bottleneck

Every robotic action must come from real-world experience — combining vision, sound, touch, balance, and motion in 3D space. Unlike digital AI, robots cannot just “simulate” experience at internet scale. Each recorded action requires physical hardware, environmental setup, and time.

• Every interaction wears down motors, joints, and batteries.

• Every experiment consumes money, materials, and human supervision.

• Every dataset is bound by the limits of time and physical space.

This is why software AI has surged ahead while embodied AI — the kind that powers robots — remains in its infancy.

Traditional Training Methods

Historically, robotics research has relied on two primary approaches: teleoperation and simulation.

Teleoperation involves human operators manually controlling robots, recording each movement as a demonstration. These recordings form the foundation for imitation learning. While accurate, the process is painfully slow, expensive, and unscalable — each dataset requires specialized hardware and skilled human operators.

Simulation allows for infinite virtual experiments, reducing cost and risk. However, the sim-to-real gap remains a fundamental barrier. Simulated data rarely transfers accurately to real-world conditions, where lighting, friction, deformable objects, and unpredictability vary endlessly.

Both methods have value but neither provides the scale or diversity required for general-purpose robotics intelligence.

The Bottleneck — and the Opportunity

As the Coatue chart illustrates, robotics today is operating at data poverty levels compared to other modalities. Language models had the internet. Vision models had billions of open datasets. Robotics has only a handful of lab-generated experiences.

This scarcity is not just a limitation — it is the core opportunity of the coming decade.

Whoever solves the robotics data bottleneck — whoever can collect millions to billions of real-world, human-level interaction datasets — will unlock the next trillion-dollar wave of AI: Physical AI.

When robots can be trained on vast datasets of human experience — cooking, cleaning, folding, driving, assembling, repairing — they will begin to learn not just patterns, but skills. They will generalize across tasks, adapt to new environments, and truly act autonomously.

The future of robotics will not be decided by better hardware or faster chips alone, but by those who can capture and align the world’s real physical data — at scale.

That is the missing layer between software AI and embodied AI — and it is the foundation on which the next generation of intelligent machines will be built.