A robot arm moving a block is a useful benchmark. A technician diagnosing a fault, a cook preparing a service, or a crew restocking a store is closer to the work models will eventually need to understand. Those tasks run for tens of minutes and change shape as they go.
Why duration is hard
Long sessions stress every part of the stack. Storage fills. Batteries drain. Clocks drift. A camera gets bumped at minute forty and no one notices until review. The longer the session, the more likely that the interesting part sits next to a failure.
Checkpoints and resumability
Recordings are written in chunks with a manifest at every boundary. If a device loses power, the session can resume without overwriting what came before. Each chunk carries an integrity hash, and duplicate detection runs before upload so a retried transfer cannot quietly double a dataset.
Time alignment
Every stream is timestamped at capture and aligned against a common clock. We measure drift during the session and correct it in post-processing rather than assuming the devices stayed in step. Reviewers see a single timeline with video, audio, and sensor traces stacked together.
Keeping the context
The pixels are only half the record. Long-horizon sessions also capture session notes, environment metadata, task boundaries, interruptions, and the operator's own account of what changed. Without that context, a model learns the shape of the room but not the reason anyone was in it.
What we are releasing
Our long-horizon program now covers tasks from ten to ninety minutes across logistics, assembly, and maintenance. Protocols are configurable, and partners can request specific task structures, camera layouts, and annotation depths. The first datasets are shipping through the API beta.