Wrist-view manipulation data,
ready to license.
Egocentric video + synchronized 6-axis IMU from a wristwatch camera —
the exact viewpoint robot wrist cameras consume. The dataset packs below are ready for robot hand learning.
Task catalog
Pick a task below to see its sample video and send an inquiry — we reply with scope, license terms, and a quote.
Pack 01
Cloth manipulation
towels · garments — folding, unfolding, smoothing
Wrist video + synchronized IMU · commercial, non-exclusive license · volume on request
Laid over the IMU trace: teal = how far the view departs from moving as one piece, shaded = its strongest stretches.
Pack 02
Overhead & fixture tasks
lightbulbs · hooks · reaching and turning overhead
Wrist video + synchronized IMU · commercial, non-exclusive license · volume on request
Laid over the IMU trace: teal = how far the view departs from moving as one piece, shaded = its strongest stretches.
Pack 03
Fastening & knots
shoelaces · zippers · buttons · cable ties
Wrist video + synchronized IMU · commercial, non-exclusive license · volume on request
Laid over the IMU trace: teal = how far the view departs from moving as one piece, shaded = its strongest stretches.
Pack 04
Containers & lids
jars · bottles · screw caps · snap lids · boxes
Wrist video + synchronized IMU · commercial, non-exclusive license · volume on request
Laid over the IMU trace: teal = how far the view departs from moving as one piece, shaded = its strongest stretches.
Pack 05
Packaging handling
bags · pouches · unwrapping · tearing open
Wrist video + synchronized IMU · commercial, non-exclusive license · volume on request
Laid over the IMU trace: teal = how far the view departs from moving as one piece, shaded = its strongest stretches.
Pack 06
Kitchen prep & clearing
setting the table · clearing · putting away (no cooking)
Wrist video + synchronized IMU · commercial, non-exclusive license · volume on request
Laid over the IMU trace: teal = how far the view departs from moving as one piece, shaded = its strongest stretches.
Pack 07
Sorting & tidying
shelving · drawer organizing · object placement
Wrist video + synchronized IMU · commercial, non-exclusive license · volume on request
Laid over the IMU trace: teal = how far the view departs from moving as one piece, shaded = its strongest stretches.
Made to order
Your task, made to order
Need a task that isn't in the catalog? Tell us the manipulation, how many episodes, and the environment — we build the dataset for you: scope and quote within 24 h, collection in 2–3 days, delivery within about a week — with the same consented participants, hardware, and format as every pack above.
Why the wrist?
Modern manipulation policies don’t watch from a distance — they read cameras mounted on the robot’s own wrists. π0, for example, takes three camera streams as input, and two of them are wrist views.
Most human video datasets are recorded from the head. A head camera sees what a person looks at — not what the hand touches; the hand–object contact region is routinely occluded. We record from a wristwatch camera —
every episode is captured from the viewpoint a robot’s wrist camera will use, with a synchronized IMU carrying the wrist’s own motion.
Does it actually help?
We ran a controlled test on a public cloth-folding benchmark. A policy was trained on a small set of robot demonstrations, and the only thing we changed was whether its vision encoder had been pretrained on this data.
81 training runs, 840 evaluation rollouts, all scored on the same unseen garments and placements. Reproduced across four pretraining variants. Stated plainly: the gain is largest where robot demonstrations are scarce, and it closed once demonstrations were plentiful — this data substitutes for demonstrations you would otherwise have to collect. Full write-up on request.
What you get
Every number below is measured on our hardware, not estimated.
| View | Wrist-mounted, palm-facing — continuous view of the hand–object contact region |
| Video | 640×480 @ 30 fps H.264 · 20-second complete task episodes (start state → end state) |
| IMU | 6-axis (ST LSM6DSO), ~100 Hz effective, ±16 g / ±2000 dps |
| Sync | Per-clip self-describing header; IMU logger leads video by ~0.6 s with the offset documented per clip |
| Integrity | 0.00% malformed rows · single-sample accel outliers ~1.4% in raw (removed in processed) |
| Format | LeRobot v3.0 + raw mp4/csv + sha256 manifest |
| Consent | Participants consented, incl. commercial redistribution. Serial-only IDs — no PII in the dataset |
FAQ
Can I train commercial models on this?
Yes. The standard license is commercial non-exclusive: train, fine-tune, evaluate, and own the resulting models. Redistribution of the raw data is not allowed.
Do episodes come with labels?
Task-level metadata ships with every pack: task id, subject id, success flag, and per-clip sensor headers. We do not provide hand-pose or action labels.
Do you collect custom tasks?
Yes — see 'Made to order' below the catalog. You specify the manipulation task and environment constraints; we collect with consented participants on our hardware.
Where does the data come from? Is any of it patient data?
No. This dataset is collected in dedicated sessions with separately consented participants. It contains no patient data, no medication records, and no data from our healthcare products.
Can I redistribute derived datasets?
Not under the standard tier — a Commercial + Redistribution tier covers derived-dataset publishing; ask us in your inquiry. Reselling the original data is never allowed.
Also from this hardware
The same wrist hardware that records our datasets now drives a duck.
Vibe-Watch × MicroDuck →