Egocentric video, hand pose, gaze tracking, and force gloves for VLA pre-training.
Human demonstration is the process of a trained operator physically performing a task — grasping, pouring, assembling, navigating — while wearable sensors capture hand pose, gaze direction, body motion, and egocentric video. The resulting dataset teaches a policy how a human actually solves the task, not how a script approximates it.
Collecting demonstration data in-house means recruiting skilled operators, sourcing sensor rigs, building a logging pipeline, and running QA across thousands of episodes. We compress that into a managed program.
Why outsource demonstrations?
Your research team should be training models, not managing capture sessions. We handle recruitment, hardware, and quality so you get clean episodes on schedule.
50,000+ hours collected to date.
40 Hz hand pose tracking across all rigs.
3 continents of operator coverage.
Where we collect
41+ delivery centers across 12 countries. Every program runs from a Roborax hub near your target time zone.
Asia Pacific
India · Philippines
Americas
USA · Canada · Colombia · Jamaica · El Salvador · Belize
EMEA
UK · Albania · Kosovo · Morocco
Egocentric capture across hands, gaze, scene, and force — synchronized into a single timeline.
First-person scene capture at 60fps, calibrated for VLA pre-training.
Per-finger joint angles plus grasp state, at 40Hz.
Operator gaze fused with scene video for attention modeling.
Glove-based force data for contact-rich tasks.
A four-stage pipeline that lands clean multi-modal data in your bucket every week.
Spec the task domain and the demonstration patterns. Coverage matrix locked.
Train demonstrators on the capture protocol. Per-task acceptance criteria reviewed.
Synchronized multi-modal recording. On-rig validation flags drift.
Annotated, time-aligned dataset delivered as bag files or your custom format.
Best-in-class hardware for each modality.
Research glasses
Hand tracking
Per-finger gloves
Gaze glasses
Egocentric video
3rd-person sync
FAQ
Teleoperation uses a control interface — the operator never touches the robot. Human demonstration captures natural human motion directly, which is then used to train imitation learning policies. Both produce trajectory data but from different sources.
Operators are briefed on task objectives but not scripted on exact movements. This produces natural variation in approach and execution — which is exactly what imitation learning needs to generalize.
Our operator network spans diverse demographics, hand sizes, grip strengths, and experience levels. If your task requires a specific operator profile — handedness, physical dimensions, domain expertise — we can filter for it.
Every demonstration is reviewed against a task specification before acceptance. Outliers are flagged and sent back for re-capture. Inter-operator consistency metrics are included in every delivery report.
From the blog
VR Teleop vs. Physical DemonstrationWhich method produces better training data and when.
From the blog
Teleop Operator Fatigue: The Hidden VariableHow fatigue affects data quality and what to do about it.
Specify the scene domain and modality mix. We scope a sized program in two days.
What actually predicts remote operator performance, selection tests that work, an onboarding curriculum, realistic ramp times, and keeping skills from decaying.
What a managed data workforce should actually include, the six questions that separate supervision from a labour pool, and why per-operator tracking matters.
What licensed corpora and bespoke capture are each good for, how to evaluate a dataset before buying, and when custom collection is unavoidable.
Why kitchens combine every hard robotics problem at once, where policies fail, what must be captured, and how to grade success when done is a judgement call.
The questions that actually predict whether a robot data partner delivers: quality measurement, schema interoperability, operations, commercial terms, and compliance.
Why humanoid datasets differ from bimanual ones, the streams they must include, where collection volume goes, and the gaps that surface in deployment.
VR and leader-follower robot control logging.
RGB-D, LiDAR, force, and tactile streams.
Bounding boxes, segmentation, action labels.
Domain-randomized scenes and sim transfers.
Held-out test sets and success-rate scoring.
Rare scenarios your policy will face in production.
