starpilot-ai/umiSample

UMI Sample

A curated sample of bimanual UMI-style gripper manipulation data recorded by Starpilot.

• 31 episodes · 1.1 hours · 119,267 frames @ 30 fps
• 7 distinct tasks across 5 activity cycles
• 12 operators · 5 recording cells · recorded 2026-07-16 to 2026-07-19
• LeRobot v3.0, robot type starpilot_yam_gripper

Every episode carries the same 6 synchronised video streams — two scene cameras (base_0, base_1), both wrist RGB cameras, and both wrist depth streams — plus a 46-D observation state and a 14-D action, at 30 fps.

How this sample was selected

Every episode here was individually inspected and confirmed by a human. A curator stepped through candidate episodes one at a time — contact sheet, both scene cameras, and full metadata — and decided each one in or out. Episodes already rejected in QA (operator error, incomplete recording, inactivity) were never offered.

How much to trust the labels

Task labels in the source recordings are inferred from an episode's position within its recorded cycle. That inference breaks whenever an operator skips or repeats a step, so the labels are not uniformly reliable, and this dataset is explicit about which are which:

• 19 of 31 episodes carry a label a reviewer checked against the footage in a separate QA pass (1 of those had the inferred label corrected).
• 12 of 31 episodes carry the positional inference, confirmed for inclusion but never separately label-verified. Treat these as weaker evidence.

Coverage

The confirmed set spans 7 tasks across 5 cycles, 12 operators and 5 recording cells. It is a sample, not a balanced corpus: candidates were worked longest-episode-first, so long-running tasks are better represented than short ones, and it does not cover every task in the underlying corpus.

Each episode was re-cut from its source recording, and any trim a reviewer set (marking where the task really starts and ends) was applied, so the footage here is the confirmed portion and nothing else.

Tasks

• C33 — Make & serve different hot drinks
• C34 — Wash up after cooking (sink-free)
• C36 — Restock the vanity
• C37 — Tidy & rearrange the office desk
• C38 — Set up & tear down the phone stand (desk B)

Privacy

Operators are visible on the scene cameras in this dataset. Operator NAMES have been replaced with stable operator_NN ids, but faces, bodies and clothing were not blurred.

Operator ids (operator_00, operator_01, …) are stable across rebuilds of this dataset but carry no meaning outside it.