Short answer
Continuous shift recordings capture everything a camera sees across hours of work, including idle time, interruptions and incidental people. Curated task clips are segmented, labeled units cut around one task attempt. Continuous footage costs more to store, review and clear for rights; clips cost more to prepare but are easier to train on, evaluate and license narrowly.
The decision behind the format
Teams often ask for "all the footage" and discover later that the format decided their budget. A single eight-hour shift from one camera is roughly 29,000 seconds of video, most of which is not a task attempt. Someone has to find the attempts, name them, check them for people and screens, and decide what a training example is. That work happens either at the supplier, before delivery, or at the buyer, after; the buyer-side version is described in curating raw workplace video, and step boundaries in step-segmentation annotation.
This page compares the two formats on the dimensions that change cost and risk. It does not decide whether to use existing footage at all; that is the question of existing workplace footage vs commissioned recording.
What each format is
Continuous recordings run for a shift, session or fixed block with no editing. They include waiting, walking, conversation, phone use, breaks and task attempts in the order they happened. They preserve context: what came before an error, how long a changeover took, how often a worker was interrupted.
Curated task clips are cut from continuous recordings or captured one attempt at a time. Each has a start and end, a task label, often step boundaries, an outcome flag and metadata. Published collections show that annotation can dominate effort even when capture is simple: Ego4D reports over 250,000 hours of annotator effort for its 3,670 hours of video (Grauman et al., 2021).
A middle option exists: deliver continuous recordings with an index of task intervals, so a buyer can use clips without losing context. EPIC-KITCHENS-100 takes this shape, with 700 variable-length videos annotated in stages starting from "pause-and-talk" narration, then transcription, parsing and temporal segmentation (Damen et al., 2020).
Format comparison
| Dimension | Continuous shift recording | Curated task clips |
|---|---|---|
| Training units | Must be derived by you or the supplier | Delivered ready to use |
| Labeling effort | Search for events first, then label | Label known units directly |
| Storage and transfer | Large; most bytes are non-task | Smaller for the same task coverage |
| Context for errors, pacing, interruptions | Preserved | Lost unless added to metadata |
| Rare events | May be present; hard to find | Present only if selected |
| Person exposure | High; incidental people for hours | Lower; edited to task windows |
| Rights and consent scope | Broad; harder to define | Narrower; easier to describe in a license |
| Risk of selection bias | Lower in capture, shifted to your curation | Higher; the curator picks what looks good |
Selection bias deserves attention. Clips cut by an annotator tend to be clean, complete successes. If you want recoveries, hesitations or mistakes, say so in the specification, as covered in failure and recovery demonstrations.
What neither format supplies
Both formats are video. Neither contains robot actions, joint states or contact forces, and both inherit the camera's viewpoint. A continuous workplace recording is still human footage; it does not become robot control data by being long. The background to this limit is in what human demonstration video can and cannot do.
Continuous recording gives you more material to mine, not more kinds of signal. If your architecture needs 3D hand pose, calibrated cameras or synchronized IMU, specify those at capture; an archive that never recorded them can at best support rough estimates.
When each format fits
- Choose continuous recordings when you are mining for rare events, studying pacing and idle behavior, building a long-horizon planning set, or pretraining on scene dynamics where task boundaries matter less. Ego-Exo4D, for example, records multi-view takes of skilled activity with expert commentary and then adds benchmark annotations (Grauman et al., 2023).
- Choose curated clips when training an imitation or co-training model that expects one attempt per example, building an evaluation set, or when your rights position requires a narrow, reviewable unit.
- Choose indexed continuous when you want clip convenience but need the surrounding context, and can afford the metadata work.
Storage is the other budget line: raw multi-hour video from several cameras adds up quickly, so read storage and transfer planning for video before asking for everything. Whichever unit you choose, each one needs a record of site, camera and consent scope, as in the demonstration clip metadata schema.
Rights and permissions consequences
The format changes what has to be cleared. This section raises diligence questions; it is not legal advice.
- Continuous recording of a work area records people for hours who may never have been part of a demonstration. Regulators publish guidance on video devices, including the EDPB's guidelines on personal data processed through video devices (EDPB, 2020). Ask the supplier how its original recording purpose and notice compare with training use.
- Incidental content such as screens, documents, customer interactions and whiteboards appears far more often in continuous footage.
- A narrow license is easier to write around clips. A license for continuous footage should say what happens to the unused portions: deleted, retained by the supplier, or delivered and restricted.
- Redaction at scale on continuous video is costly, and blurring faces does not by itself resolve permission questions. See redaction in demonstration video and worker consent and identifiable people.
For new recordings, consent can be designed in for either format, as discussed in recruiting and briefing demonstrators.
Illustrative delivery specification
# Illustrative only - synthetic values
delivery_unit: indexed_continuous # or: task_clip
recordings:
duration_per_session: 4 h
container: mp4 (H.264), constant frame rate, timestamps preserved
index:
task_intervals: [start_s, end_s, task_id, outcome]
non_task_intervals: labeled (idle, walking, conversation)
interruption_flags: yes
clips_derived:
padding_before_after: 2 s
min_length: set by buyer
retention:
unused_portions: deleted by supplier within period agreed in license
metadata: [site_017, operator_A12 (pseudonymous), camera_mount, session_id]
Questions to ask suppliers
- Are recordings continuous, edited or both, and can I see an unedited sample from one session?
- What share of recorded time is task activity?
- Who segmented the clips, with what instructions, and what was discarded?
- How are incidental people, screens and documents handled in continuous footage?
- What notice and consent covered the people in unused portions?
- What happens to non-delivered footage under the license?
- What storage and transfer volume should I expect per task-hour?
- Can you deliver an interval index so I can re-cut clips myself?
Discuss your requirements
If you are choosing between continuous capture and task clips, describe the format, tasks and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery; where no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match. Return to the humanoid training data guide or the workplace task demonstration hub, or browse the catalog page on video recordings.
Describe the workplace demonstration data your robotics project needs
Sources
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — corpus scale, consent, de-identification and annotation effort.
- Rescaling Egocentric Vision — variable-length videos with staged narration-based annotation.
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — long multi-view takes with expert commentary.
- Guidelines 3/2019 on processing of personal data through video devices — regulator guidance on video devices.