Pantheon releases DiverseUMI-100 with 12,219 robot tasks

Pantheon has released DiverseUMI-100, a public sample of its larger robot-manipulation dataset.

What happened

The release contains 99.6 hours of bimanual tabletop manipulation across 12,219 task episodes and 1,134 recordings. Each task includes synchronized overhead and wrist-camera video, wrist pose and action data, and detailed labels describing the activity and outcome.

The dataset is publicly accessible after accepting a custom commercial-use licence. That licence permits research, model training and commercial use, while restricting resale or redistribution of the underlying data.

Why it matters

General-purpose robots need datasets that cover many objects, environments and actions rather than thousands of repetitions of one narrow task.

A large, densely labelled manipulation dataset gives developers more material for training and evaluating models across varied physical behaviours.

The bigger picture

Robot data is beginning to resemble core AI infrastructure. Dataset size matters, but diversity, labelling quality and clear usage rights will determine how useful each release is.