origin-data-lab/egocentric-keystep-annotation

Keystep Annotation and IMU Boundary Evaluation

Origin Data Lab · Technical Field Note #05 · 24 October 2026

Short answer. Structural checks found zero reported gaps, overlaps, missing required fields or out-of-vocabulary verbs across 25 human-labelled keysteps in two egocentric cooking clips. A body-worn IMU showed measurable motion changes at some boundaries, but not all task-to-task boundaries were distinguishable from random positions. IMU motion is a useful review signal, not proof of fine hand-action boundary accuracy.

Full article: Technical Field Note #05 — Keystep Annotation for Egocentric Robot-Learning Data

Code and supporting evidence: GitHub — origin-data-lab/egocentric-keystep-annotation · Hugging Face files

---

Key results

Metric	Sample A — knife preparation	Sample B — seasonings and stirring
Clip duration	66.0 s	65.967 s
Frames (30 fps)	1,979	1,978
Human-labelled keysteps / phases	13 / 3	12 / 3
Gaps / overlaps	0 / 0	0 / 0
Missing required fields	0	0
Out-of-vocabulary verbs	0	0
Reported IMU-to-video alignment offset	188 ms	200 ms

Native IMU sampling rate: 410.6 Hz. All findings are from two clips from one operator, not a population-level benchmark.

Human-labelled spans alongside the measured gyroscope signal; labels and motion do not always align one-to-one.

What the boundary test shows

The test compares motion changes around labelled keystep boundaries against 2,000 random boundary sets per comparison. The motion score uses gyroscope magnitude in windows around each boundary. Two comparisons include all boundaries; two exclude transitions involving non-task segments.

Boundary set	n	Labelled score (°/s)	Random mean	Random p95	p
A · all	12	21.9	7.2	14.7	0.005
B · all	11	29.9	10.0	20.1	0.002
A · task-to-task	8	3.3	4.1	6.8	0.67
B · task-to-task	7	10.2	3.9	7.5	0.003

Interpretation. Sample A's task-to-task boundaries were not distinguished from random positions by this test (p=0.67), while Sample B's were (p=0.003). Strong coarse-motion changes should not be mistaken for reliable detection of every hand-action transition.

Evaluation files

Path	Contents
boundary_motion_results.csv	Four observed-versus-random boundary comparisons; suitable for tabular inspection
annotation_example.jsonl	One documented keystep annotation (ks_A_01, 0.0–7.5 s)
figures/	Five figures from the technical note, including the timeline and boundary test

The public release is a limited evaluation sample. It does not contain the original videos, raw IMU signals or complete annotation files for all 25 keysteps.

Annotation format

The JSONL example describes an observed action span using id, t_start, t_end, phase, verb, object, tool, active_hand, support_hand, instruction, state_before, state_after and outcome.

Representative frames illustrate the action sequence; they are not a replacement for source video or full annotations.

What this evaluation does not verify

• Semantic correctness of all labels. Structural QA verifies completeness and consistency, not whether each action was labelled correctly.
• Fine-grained hand-action boundaries. A wearer-mounted IMU may miss subtle hand manipulations or register unrelated torso/head movement.
• Generalization. Two clips from one operator cannot establish reliability across people, tasks, devices or environments.
• A complete training dataset. This release contains one annotation example and four boundary-test rows, not 25 downloadable keystep records.

Cite

@misc{origindatalab2026fieldnote05,
  author       = {{Origin Data Lab}},
  title        = {Keystep Annotation and IMU Boundary Evaluation},
  howpublished = {Technical Field Note \#05},
  year         = {2026},
  url          = {https://origindatalab.io/technical-field-notes/keystep-annotation-egocentric-robot-learning.html}
}

Or:

Origin Data Lab (2026). Keystep Annotation and IMU Boundary Evaluation. Technical Field Note #05.
https://origindatalab.io/technical-field-notes/keystep-annotation-egocentric-robot-learning.html

License

This evaluation package is marked rights reserved. It is not licensed under the CC BY-NC 4.0 terms used by some earlier field-note repositories. Request permission before redistributing or reusing the text, results or figures beyond applicable exceptions.

Related

• All Technical Field Notes
• 한국어 — 키스텝 어노테이션 검증 (Field Note #05)
• #04 Stereo Calibration Validation · dataset card

---

Origin Data Lab — real-world data production for Physical AI and robotics · origindatalab.io · contact@origindatalab.io

Need evaluation-ready egocentric data with annotation and QC requirements tailored to your project? Contact us.