AI & Humanoid Robotics Weekly

Robot training data

Companies that collect, label and sell the real-world data robot AI models learn from: human motion and video, teleoperation and robot-run logs.

Data updated

Why it matters

Robot foundation models cannot be trained on internet text the way language models are, so they need large amounts of physical-world data. Suppliers pay people to record everyday tasks with body sensors, head cameras or phones, or run fleets of teleoperated robots, and sell the datasets, custom collection and model evaluation to robot makers and AI labs.

Why it is a bottleneck. Data is widely seen as the main limit on general-purpose robot models. It is slow and costly to collect, and it must match the robot's body and tasks. Scale AI says the main open datasets, such as DROID and Open X-Embodiment, hold only about 5,000 hours of robot interaction data combined; suppliers now sell data by the hundreds of thousands of hours.

Cost. Collection labor can be cheap: Human Archive pays collectors in India a base rate of $1 an hour (TechCrunch, May 2026). Much of the cost is in capture hardware and quality review: Maxinsights says gloves on the market yield only 10 to 20 minutes of usable capture a day, and that its review software cut review costs by about 80%.

Suppliers 4

  • MaxinsightsFirst-person video of real work from paid contractors; quality-scoring software (MaxLoop); a touch and force capture glove (MaxGlove, developed with Genesis AI)
  • Mecka AIFirst-person human motion datasets (body sensors, iPhones), custom data collection and model evaluation
  • Scale AICustom robot and human-demonstration data collection and labeling; data-capture software in Universal Robots' AI Trainer
  • XDOFTeleoperated robot data, wearable-sensor human data and annotation; ABC-130K open manipulation dataset

Sources for this category