Short answer
Demonstration pace is a specification choice, not a detail. Production-speed recordings show how work really looks and flows but blur hands and compress motions; slowed recordings improve visibility and tracking but depart from realistic timing. For most buyers the safest specification records both on matched tasks, tags every clip with its pace, and states the capture settings that keep fast motion legible.
Why pace changes what the data is worth
Three things move with pace: what the camera can resolve, how realistic the behavior is, and how much a model can learn from a single pass.
Visibility. A hand moving quickly leaves motion blur at long exposure times and gaps between frames at low frame rates. Both degrade hand tracking, contact timing and any label that depends on seeing fingers. A demonstrator who slows down gives annotators and pose estimators more frames per movement.
Realism. People who are asked to demonstrate often change how they work: they pause, exaggerate, narrate, check the camera. The result is cleaner but not what the job looks like. Production pace preserves natural shortcuts, overlaps between steps and the micro-recoveries that real work contains.
Learnability. A policy has to produce actions at its own control rate, which depends on the robot and architecture. DROID, for example, ran at 15 Hz control on its teleoperated arm (Khazatsky et al., 2024). Human motion at production pace and a robot's achievable speed are different things, and neither fixes the other. The papers cited on this page do not isolate pace as an experimental variable, so treat what follows as procurement reasoning and test it on your own pilot.
What published work does and does not tell you
For the wider limits of video as a source, see what human demonstration video can and cannot do. Published imitation results are about how well methods use whatever demonstrations they are given. ACT reports 80-90% success on 6 real fine-grained tasks from about 10 minutes of demonstrations (Zhao et al., 2023). Diffusion Policy reports a 46.9% average improvement over prior approaches across 12 tasks from 4 benchmarks (Chi et al., 2023). Both are results on robot demonstrations, so they do not settle how human video should be paced.
For human video, Ego-Exo4D is useful because it records skilled activities from first-person and several third-person cameras at once, with expert commentary and a proficiency estimation benchmark (Grauman et al., 2023). It shows that skill level is itself a labeled property of a recording, which is the right way to handle pace: record it as metadata rather than assuming it away. For the related question of demonstrator skill, see the expert vs novice demonstrations page.
Decision table
| Option | Strength | Weakness | Fits when |
|---|---|---|---|
| Production pace, unmodified | Realistic flow, true durations | Blur, occlusion, missed frames | Task understanding, progress estimation, planning |
| Slowed, demonstration-style | Clear hands, easy labeling | Unnatural timing, pauses | Hand tracking, fine-grained segmentation |
| Paired: slow pass then normal pass | Both views of one task instance | Doubles recording time per task | Pilot studies, label bootstrapping |
| Production pace with high frame rate | Realism plus legible motion | Larger files, more lighting needed | Co-training with tracked hands |
| Time-stretched afterward | Cheap | Invents no new information; blur remains | Rarely a substitute for capture |
Post-hoc slowing is worth singling out. Slowing a clip in editing does not add detail that the original frames never captured. If you need legible fast motion, buy it at capture time through shorter exposure, higher frame rate and adequate lighting. See head-mounted camera capture specification for device settings.
Illustrative pace specification
# Illustrative only - synthetic values, set your own from a pilot
pace_protocol:
tags: [production, instructed_slow]
production:
instruction: "Work at your normal shift pace; do not narrate or look at camera"
frame_rate: 60 fps (30 fps acceptable if exposure under 1/250 s)
instructed_slow:
instruction: "Same task, deliberate pace, hands in frame, pause 1 s at each step end"
pairing: each demonstrator records production first, then instructed_slow, same session
metadata_per_clip: [pace_tag, operator_id, session_id, task_instance_id, fps, shutter]
acceptance:
hand_visible_frames_min: 85% # buyer-set
blur_check: sampled frames reviewed for finger edge definition
pause_ratio_max: 10% # production tag only
Briefing and verifying pace on site
Pace instructions fail when they are vague. "Work normally" gets interpreted differently by each demonstrator, and "go slowly" produces anything from a smooth half-speed to a sequence of freezes. Write the instruction as observable behavior: hands stay in the camera's field of view, no talking during production passes, a one-second hold at each step end during slowed passes. Have the collection provider run two or three rehearsal tasks and review them with you before the main session.
Verification should not rely on demonstrators' own sense of speed. Compute task duration per instance and compare the production and slowed passes: a slowed pass that is only 10% longer than production probably did not change much, while one that is four times longer probably contains idle time. Look at hand speed on a few annotated segments, and spot-check the frames at the fastest moments for finger edge definition. Record these checks in the delivery notes so a later reader knows the data was screened for blur and not just for length.
Finally, keep pace separate from skill. A fast expert and a fast novice are different data. If the pilot shows demonstrators drifting toward faster, sloppier work over a long session, shorten sessions or add rest rather than correcting afterward.
Limits and extra preparation
- Pace mismatch with the robot. A humanoid's actuators, safety limits and balance control set its achievable speed. Retargeting from a person works with trajectories, so timing usually has to be rescaled or discarded. See the human-to-humanoid embodiment gap.
- Mixed pace in one training set. Without a pace tag, a model learns an average. With the tag, you can condition, filter or weight.
- Frame-rate hazards. Variable frame rate video and dropped frames corrupt timing. Ask for timestamps per frame, not only a nominal rate.
- Instructed slowness is an act. Check whether demonstrators slowed by moving smoothly or by inserting pauses; those are different distributions.
- Safety. Do not ask workers to run faster than their normal practice for the camera. Production pace means their normal pace.
Questions to ask collection providers
- How will demonstrators be briefed on pace, and is the wording in the protocol?
- Can you record matched production and slowed passes of the same task instance?
- What frame rate, shutter or exposure limit, and lighting do you commit to?
- Are timestamps stored per frame? How are dropped frames flagged?
- How do you detect demonstrators who change behavior because they are being filmed?
- Will the pace tag, operator and session identifiers be delivered with each clip?
- What review is done on motion blur before acceptance?
- Does speed vary by shift or fatigue, and are shifts recorded?
Review the pace tag alongside the acceptance rules in demonstration video acceptance criteria.
Discuss your requirements
If you are specifying a commissioned recording and need to fix pace, frame rate and viewpoint, describe the tasks and capture constraints to SourceX. SourceX describes a separate physical-world path for new recordings of hands-on work, where each opportunity has its own requirements. It reviews your requirement and follows up about potential sources; a request does not guarantee a match. More context sits in the demonstration data hub and the humanoid training data guide.
Discuss your humanoid training-data requirements with SourceX
Sources
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware — ACT results on fine-grained tasks from limited demonstrations.
- Diffusion Policy: Visuomotor Policy Learning via Action Diffusion — Diffusion Policy reported improvement across benchmark tasks.
- Ego-Exo4D: Understanding Skilled Human Activity from First- and Third-Person Perspectives — Ego-Exo4D multi-view skilled-activity recording with expert commentary.
- DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset — DROID scale, rig, 15 Hz control and per-episode calibration and language.