Technical Field Note #28 · November 2026

Usable hours are not usable clips: how many 10-second clips an hour of real work gives

Field Note #28 · Usable clip length for egocentric stereo+IMU capture (ODL-EGO S1), 10 working days

Field Note #27 counted usable seconds over ten working days (20.6 hours). Robot-learning buyers usually need continuous clips, so here we measured how those seconds join up. On the five bare-hand days, 51.9% of seconds were usable, but the median unbroken usable stretch was 2 seconds. Counted strictly, an hour of recording held 42 usable clips of at least 10 seconds and 1.8 of at least 30 seconds; only 19% of recorded time sat inside a 10-second-plus clip. Most breaks were single seconds: of the isolated one-second gaps inside usable stretches, about half failed only the hand check, and in 24 of those checked by eye the hand was still visible in 22. Allowing one failed second inside a clip raised the yield to 66 ten-second clips per hour and 37% of recorded time. On glove days the same counts were 8.5 and 16.4. Delivery quotes should therefore state clip length and gap tolerance, not only usable hours.

Published November 14, 2026⏱ 11 min readStatus Preliminary · automated analysisOpen data GitHub · Hugging Face · Kaggle
42 strict · 66 with 1-s gap10-s+ usable clips per recorded hour (bare hands)5 days, 10.35 hVerified (automated)
2 sMedian unbroken usable stretchbare-hand days; 51.9% of seconds usableVerified (automated)
19% strict · 37% with 1-s gapRecorded time inside 10-s+ clipsbare-hand daysVerified (automated)
22 / 24Hand still visible in hand-only gapsrandom 1-s gaps, checked by eyeReviewed (small sample)

Why this matters

Buyers train on clips, not on seconds

A usable-hours figure counts seconds wherever they fall. A manipulation policy is trained on continuous clips: a grasp, a pour, a tie, usually several seconds long. Fifty percent usable seconds scattered in one-second pieces is worth far less than the same share in long runs.

Field Note #27 gave per-second results for ten working days. This note asks the buyer's next question: how many clips of a given length does one recorded hour produce?

Setup

Same per-second index, joined into runs

Data: the Field Note #27 index, one row per second for 20.6 hours (74,044 s) of head-mounted produce-portioning work, 2026-11-01 to 11-12, one R8 stereo+IMU rig (ODL-EGO S1). Five days in work gloves, five bare-handed.

A second is usable when exposure is OK, camera blur ≤ 8 px, the wearer looks at least 30° down and a hand is found (as in Field Notes #25 and #27). Consecutive usable seconds within a recording form a run; consecutive 5-minute files of one session are joined. "Strict" ends a run at the first failed second; "1-s gap allowed" keeps a run going across one failed second, but not two.

To see why runs break, we took the isolated one-second gaps (usable before and after) and listed which check failed. From the bare-hand gaps that failed only the hand check, 24 were drawn at random and checked by eye.

Finding 1

Half the seconds are usable, but they come in short pieces

On bare-hand days 51.9% of seconds were usable, yet the median unbroken usable stretch was 2 seconds and the 90th percentile 9 seconds. Counted strictly, one recorded hour held 42 clips of at least 10 seconds and 1.8 of at least 30 seconds. Only 19% of recorded time sat inside a clip of 10 seconds or more.

Glove days were far lower at every length (8.5 ten-second clips per hour), because the hand check failed twice as often there (Field Note #27).

Log-scale line chart of usable clips per recorded hour against minimum clip length, for bare hands and work gloves, strict and with a 1-second gap allowed
Figure 1. Clips per recorded hour at or above each length.
Table 1. Half the seconds are usable, but they come in short pieces
Bare hands (5 days)Work gloves (5 days)
Usable seconds51.9%23.7%
Median usable stretch, strict2 s1 s
10-s+ clips per recorded hour, strict42.48.5
10-s+ clips per hour, 1-s gap allowed66.416.4
30-s+ clips per hour, strict / 1-s gap1.8 / 10.70.5 / 0.8
Share of recorded time inside 10-s+ clips, strict19.2%3.6%
Same, 1-s gap allowed37.1%7.1%

Finding 2

Most breaks are one second long, and many are detector misses

A typical bare-hand recording shows long usable stretches cut by single seconds (figure). Of the isolated one-second gaps on bare-hand days, 52% failed only the hand check and 48% failed the blur check (a quick head turn); exposure and view almost never broke a run.

Checked by eye, the hand was still visible in 22 of 24 randomly chosen hand-only gaps. These breaks come from the detector, not from the work.

Timeline of 300 seconds of a bare-hand recording: long blue usable stretches interrupted by single pink seconds (no hand detected) and grey seconds (blur)
Figure 2. One 5-minute bare-hand recording, second by second.

Finding 3

Allowing one failed second changes the yield by half

Keeping a clip going across a single failed second raised bare-hand yield from 42 to 66 ten-second clips per hour, from 1.8 to 10.7 thirty-second clips, and the share of recorded time inside 10-second-plus clips from 19% to 37%. Every bare-hand day stayed in a narrow band: 34–49 clips per hour strict and 62–74 with the gap allowed.

Whether a one-second gap is acceptable depends on the buyer: a detector miss with the hand in view is usually harmless; a blur spike may not be.

Bar chart of 10-second usable clips per hour for each of ten days, strict and with a 1-second gap: 4–13 on glove days, 34–49 strict and 62–74 with the gap on bare-hand days
Figure 3. 10-second-plus usable clips per recorded hour, by day.
Table 2. Allowing one failed second changes the yield by half
DateHands10-s+ clips/h, strict10-s+ clips/h, 1-s gap allowed
11-01gloves7.012.8
11-02gloves4.19.3
11-03gloves10.316.8
11-04gloves9.221.1
11-05gloves13.325.2
11-08bare48.864.5
11-09bare34.161.8
11-10bare43.074.1
11-11bare45.569.7
11-12bare40.563.7

What we changed

Clip yield is now part of every quote

Usable time is now reported together with clip yield:

  • Each delivery states clips per hour at the buyer's minimum length, with the gap rule used (strict, or one failed second allowed and why it failed).
  • Bridged seconds are flagged in the index, so a buyer can drop them or keep them.
  • Quotes for robot-learning buyers start from clips of the required length, not from recorded or usable hours.

What this means

Ask for clips at your length

Verified: on bare-hand days at this site, 42 clips of 10 s or more per recorded hour strict, 66 with one failed second allowed; stable from day to day.

Reviewed on a small sample: most single-second breaks with no hand detected still showed the hand.

For buyers: usable hours alone can overstate clip yield by a factor of two or more. Ask for clips per hour at your minimum length and the gap rule behind them.

SCOPE AND LIMITS

Limitations

What this note does not show, stated up front.

  • One site, one wearer, one rig; clip yield depends strongly on the task. Produce portioning has many short actions.
  • Per-second checks use the keyframe that opens each second (Field Note #27); a run can hide a failed frame inside a passing second.
  • The by-eye review covers 24 gaps from bare-hand days only.
  • Preliminary analysis: AI-assisted automated analysis plus a small manual review; no third-party review.

FAQ

Questions data buyers ask

Q01How many 10-second usable clips does one hour of egocentric work video give?
At this produce-portioning site with bare hands, 42 clips of at least 10 seconds per recorded hour when any failed second ends a clip, and 66 when one failed second is allowed inside a clip. With work gloves, 8.5 and 16.4.
Q02Why is clip yield lower than usable time suggests?
Usable seconds come in short pieces: on bare-hand days the median unbroken stretch was 2 seconds even though 51.9% of seconds were usable. Single-second breaks, mostly hand-detector misses and quick head turns, split longer actions.
Q03What is a 1-second gap tolerance?
A clip continues across one failed second but ends at two in a row. Here it raised the share of recorded time in 10-second-plus clips from 19% to 37% on bare-hand days.

OPEN DATA AND CODE

Reproduce and inspect the evidence

Every number in this note can be checked: derived tables, analysis code, figures and a runnable notebook are public.

Run the notebook on Kaggle

Licence: CC BY 4.0 (data and figures) · MIT (code)

Cite this note

Origin Data Lab (2026). Technical Field Note #28: Usable hours are not usable clips. https://origindatalab.io/technical-field-notes/usable-clip-length-egocentric-work-ten-days.html

Want this measured on your data?

Origin Data Lab plans and runs egocentric, stereo+IMU and field video capture with documented QA, for physical AI and robotics teams.