Training physical AI systems requires examples of successful physical behavior.

One obvious approach is teleoperation: put a human behind a robotic system and have them control the robot while its movements are recorded.

Teleoperation works. But it introduces a fundamental limitation.

The data collection process becomes constrained by the robot.

The Hardware Bottleneck

Robots are expensive, difficult to deploy, and limited in number.

If every training example requires access to a robot, collecting millions of diverse demonstrations becomes difficult. The process requires robotic hardware, controlled environments, operators, maintenance, and specialized infrastructure.

Humans already possess something robots are still trying to learn:

general-purpose physical capability.

People naturally understand how to grasp objects, adapt their movements, respond to unexpected changes, and complete tasks using whatever tools and environments are available.

Instead of asking humans to operate robots, we can capture humans performing the task directly.

Capturing the Target Behavior

A person preparing food, organizing a workspace, assembling an object, opening a container, or using a tool is already demonstrating the behavior a robot ultimately needs to reproduce.

The challenge is converting that behavior into machine-readable data.

A wearable capture system can record:

• Egocentric visual information
• Head movement and pose
• Hand and finger articulation
• Upper-body motion
• Object interaction
• 3D trajectories
• Temporal task sequences

The hardware can remain minimal, lightweight, extremely portable, and easy to deploy, while electromagnetic tracking and other sensing technologies preserve the motion signals required for downstream learning.

Humans as the Data Collection Platform

This changes the economics of robotics data collection.

Instead of building a robotic setup for every demonstration, the same capture infrastructure can travel with the person.

That means data can be collected in:

• Homes
• Kitchens
• Workshops
• Warehouses
• Offices
• Retail environments
• Industrial spaces
• Other everyday environments

The result is a much broader distribution of real-world behavior.

Better Data, Broader Distribution

Teleoperation is valuable for understanding robot-specific control.

Human instrumentation is valuable for understanding human physical behavior.

For many learning problems, the latter provides a richer starting point because humans naturally adapt to different environments, objects, and constraints.

The objective is not to replace robots in the data loop.

It is to remove robots as the primary bottleneck to collecting physical intelligence data.