01
Kitchen work
Commercial kitchens & food service
Cutting, portioning, dish handling and work-surface cleanup.
TOOL–OBJECT INTERACTIONHUMAN DEMONSTRATION DATA / INDONESIA
Human skill.
To your
specification.
Egocentric human demonstrations for physical AI. Scope task coverage, action taxonomy and acceptance criteria before collection.
Video. Temporal labels. Traceable provenance.
HOLOASSIST / EPISODE EXCERPTREAL RECORDING
HoloAssist source footage
0.00 / 38s
ACTION AT THIS MOMENT
No fine action annotated at this moment
38sone continuous task16timed action labelsMP4 + JSONsame files to import
Inspect the video, labels and loaders Scroll to continue ↓What do you need to collect?
Tell us the tasks, capture requirements and labels your model needs.
DESIGNED AROUND YOUR BRIEFTask-specific captureStructured deliveryDocumented review
01 / COLLECTION SCOPE
Start with the work.
Build the collection around it.
For teams building vision-language-action (VLA) models and physical AI. Start with the tasks, then choose the capture and supervision your pipeline needs.
Collection scenarios to scope in Indonesia. Reference photos illustrate the work; they are not Handtrained recordings or confirmed collection sites.
02
Picking & sorting
Stockrooms, logistics & workstations
Selecting parts, sorting objects, kitting and packing.
OBJECT VARIATION
03
Textile handling
Laundry, linen & housekeeping workflows
Smoothing, folding, stacking and packing fabric.
DEFORMABLE OBJECTS
04
Skilled repair
Workshops, maintenance & assembly
Fastener handling, tool alignment and ordered repair steps.
FINE MANIPULATIONImage credits & sources
Photographs are used as website illustrations only and are excluded from the downloadable dataset. All images are resized and may be cropped responsively for display.
- Chef chopping food on a red cutting board — Kamaruld Salleh. Unsplash License · Malaysia (listed on source page)
- Worker organizes metal hardware in a warehouse — Nenad Stojković (Shixart1985). CC BY 2.0. Location unverified.
- A Person Folding a Yellow Shirt on a Wooden table — Kampus Production. Pexels License · Portugal (listed on source page)
- Hands repairing a car door latch using a screwdriver in a workshop setting — Nenad Stojković (Shixart1985). CC BY 2.0. Location unverified.
CAPTURE / FIVE CONFIGURATIONSSpecify the signal.
Specify the signal.
Then choose the rig.
Start with structured manipulation. Add views, sensors or robot-aligned capture where your training objective calls for them.
Video + temporal labels + optional sensor streams
First-person RGB video with task and episode metadata for visual pretraining and activity recognition.
Reference hardwareGoPro-class camera or buyer-approved phone rig.
- Confirm before capture
- Hand visibility, field of view, frame rate and motion blur.
Video + temporal labels + optional sensor streams
Video with task → subtask → action labels, temporal boundaries and object vocabulary. Scope synchronized inertial sensors (IMU) or narration where useful.
Reference hardwareA camera with accessible IMU data, or a synchronized wearable sensor.
- Confirm before capture
- Task taxonomy, clock alignment, sensor rates and units. IMU signals do not directly measure position.
First-person and external cameras with timestamp synchronization and calibration. Add views to expose hand, object and body context.
Reference hardwareEgo-Exo4D uses Aria glasses and fixed GoPro cameras; other rigs need their own validation.
- Confirm before capture
- Clock drift, overlapping views and camera intrinsics/extrinsics. Aria research access is separate.
Scope hand tracking, depth or 3D pose with appropriate equipment and validation. Measured signals and model-derived estimates are identified separately.
Reference hardwareVision Pro / ARKit as in EgoDex; gloves or RGB-D are separate capture choices.
- Confirm before capture
- Coordinate frames, tracking loss and export access. Pose estimates do not measure contact force.
Scope a gripper or wrist-camera setup with 6-DoF trajectories and action state. Hardware, coordinate frames and transfer requirements must be validated with your team.
Reference hardwareUMI parallel-jaw grippers with wrist-mounted GoPros, or a buyer-approved rig.
- Confirm before capture
- Trajectory quality, observation/action timing and target-robot compatibility.
Illustrations show proposed configurations, not an owned equipment fleet. Device selection, procurement, site access and any research-access restrictions are confirmed in the brief. Buyer-supplied rigs can be scoped.
02 / A CLEAR WAY OF WORKING
Define the standard.
Then collect against it.
Define the brief.
Set task coverage, action taxonomy, capture modalities and the acceptance rubric around your training objective.
Calibrate together.
Review a small batch for capture quality, temporal labels and loader compatibility. Resolve gaps before scaling.
Collect. Review. Refine.
Check task completion, visibility, synchronization, duplicates and label consistency. Record exclusions and rework.
Deliver with context.
Receive versioned episodes, annotations, provenance and QA results in the agreed schema. Run your acceptance checks.
03 / BEGIN WITH A SCOPED PILOT
Small enough to assess.
Specific enough to learn.
Start with a proposed 50 accepted-hour pilot. Agree the task mix, capture protocol, license and acceptance rubric, then approve a calibration batch before the main collection.
Accepted hours count unique activity, not extra camera views. Feasibility, schedule and pricing are confirmed with your brief.
Scope your pilot50ACCEPTED ACTIVITY-HOURS
- Buyer-defined tasks and requirements
- Taxonomy, temporal labels and episode metadata
- QA report, provenance and loader validation
- Scope and pricing confirmed together
BUILT IN INDONESIA
Handtrained is an early-stage collection initiative connecting skilled people and real workplaces with physical AI teams. Participation, site permissions and intended uses are agreed before capture. Capacity and the contracting entity are confirmed for each engagement.
04 / THE PRACTICAL QUESTIONS
Clarity before
commitment.
Everything starts with an agreed scope.
01Can we try the data in our own tools?
The HoloAssist demonstration includes a Hugging Face Datasets notebook and a native FiftyOne import, tested locally on macOS arm64 with Python 3.12. It is a source-data import example, not an original Handtrained collection. For a new collection, we validate a calibration batch in your agreed loader before scaling.
02Can you work with our existing collection protocol?
Share your requirements in the collection brief. We’ll assess the tasks, hardware, site access and acceptance criteria before confirming what can be delivered.
03What counts as an accepted hour?
One hour of unique activity that meets the agreed criteria. Simultaneous camera views do not automatically multiply the billable hours. The definition is recorded in the pilot scope.
04Which formats and capture signals can be scoped?
The proposed base is first-person video with task metadata and agreed annotations. Additional sensors, synchronized views and conversions such as LeRobot or MCAP are scoped separately and confirmed before collection.
05How are permissions and usage rights handled?
Participant and site permissions must cover the proposed collection and agreed use. The project scope defines the license, delivery documentation and any restrictions. Identifying personal records are kept separate from the dataset.
06What happens if the data misses a criterion?
The pilot agreement defines the review window, acceptance rubric and rework or replacement process. These are agreed before recording starts.
YOUR NEXT COLLECTION
Discuss your
data needs.
Bring a brief. Or start with a task. We’ll define the next step together.