The robotics data bottleneck

Teleoperation records a robot performing the task, first-person video captures how people move through the same work, and simulation creates controlled episodes that can be repeated. Each source answers a different question and arrives with different sensors, timing and context.

Large language models

The Transformer became the standard

It created a shared architecture that scaled across tasks and powered ChatGPT.

ChatGPT took the breakthrough to the mass market

Robotics researchers

Researchers are testing three possible paths in parallel

Teams train each approach on different robots, tasks and datasets, then test whether it can generalise beyond its original setup, but results remain difficult to compare across labs.

• VLA / VLMVision and language to action
• World modelsLearn physical dynamics
• Sim-to-realSimulate, then transfer

Type of data	Open benchmark	How THOR scales this
Teleoperation	DROID76K demonstrations	Source robot demonstrations through the subnet across more robots, tasks and environments.
Video	Ego4D3,670 hours	Source egocentric footage from phones, glasses and wearable capture hardware.
Simulation	ProcTHOR10K interactive homes	Expand coverage with procedurally generated environments and long-tail scenarios.

• TeleoperationDROID · 76K demonstrationsSource robot demonstrations through the subnet across more robots, tasks and environments.
• VideoEgo4D · 3,670 hoursSource egocentric footage from phones, glasses and wearable capture hardware.
• SimulationProcTHOR · 10K interactive homesExpand coverage with procedurally generated environments and long-tail scenarios.

• THOR expands robot data supply through the subnet
• THOR improves robot data quality through annotation

Language models absorb trillions of tokens and image models draw from billions of paired examples. Robotics data is still counted in trajectories, many of which require a person, a machine or a useful simulation to perform the underlying action.

A model cannot use those sources well until the task has been defined, the behaviour has been described and the output has been checked. Useful supply is measured by whether that experience can train or test a model, while hours of footage alone say little.

Supply has to grow without lowering the standard. That means making physical work easier to commission and giving buyers a reliable way to reject data that does not match the intended use.