Objects and labels
Preview object detections and inferred scene or activity labels. Labels need review and do not establish what happened outside the sampled frames.
Cupel / Robotics dataReview the data
Review the data
your robot learns from.
Cupel: Public preview
Cupel offers preview enrichment for egocentric and robotics video. Sign up to try it, then review inferred outputs before using them for training. Accuracy and full-video coverage are not independently established.
Current public offerPreview outputs.
Preview outputs.
Review required.
Availability depends on the selected processing tier and source video. Preview means an output can be produced but its accuracy has not been independently qualified. Check Cupel's current product guide and pricing before starting a job.
Masks and tracks
Preview segmentation and tracking outputs are offered in Dense 2D. Tracking accuracy is unqualified; occlusion and missing frames can break a track.
Hands on sampled frames
Preview hand boxes and 2D keypoints on sampled frames. Full-video coverage is not established. 3D hand pose is unavailable; hand boxes are not yet included in frame tables or dataset formats.
Camera poses and depth
Geometry outputs require a job quote. Camera-pose accuracy and metric depth scale are not established. These estimates do not establish robot actions or a physically accurate scene.
Face blur on request
Ask through the Cupel assistant for a separate blurred copy of your uploaded video. It is not an automatic output of a processing run. Review the copy before sharing it; do not assume all identifying information is removed.
Training use
Check the delivered fields, source-frame coverage, and uncertainty against your training requirements. Preview enrichment alone does not establish synchronized robot actions, training value, or deployment readiness.
Sample and pricingTry a sample.
Try a sample.
Check the cost first.
Cupel currently offers the first 3 clips, up to 2 minutes each, free. Paid processing depends on the selected tier and source duration; geometry is quoted per job. Review the price in Cupel before processing.
Example journey / Product directionA person demonstrates.
A person demonstrates.
A robot needs to learn.
Suppose you have first-person recordings of someone sorting objects and want to prepare demonstrations for a robot arm. This example explains the intended workflow; it is not a customer result or a claim that every step is available in the public preview.
- 01
Bring the source and the target.
Start with the recordings and task instructions. Define the robot, its action space, and the training setup. Include calibration or other measurements you have, so later steps can account for what the source actually establishes.
- 02
Recover the useful parts of the demonstration.
Inspect task segments, hand motion, object tracks, and available scene information. Review uncertain or occluded moments before carrying them into a derived trajectory.
- 03
Review a robot motion candidate.
Adapt the demonstration to the selected robot. Inspect the mapping, coordinate assumptions, and relevant constraints. A plausible pose is one piece of evidence; completing the task requires its own evaluation.
- 04
Prepare a training version.
Choose examples and transformations, preserve related episodes within the same split, and export the required observations and actions. Keep the source and processing history attached.
- 05
Ask whether it helped.
Compare policy training outcomes under matching conditions. If the change does not help, use the evidence to revise processing or specify the next examples to collect.
What you take forward
A versioned dataset candidate, its source history, the evaluation evidence, and a concrete next data decision.
Before you start
Know the limits.
The public preview and the broader product direction have different evidence behind them.
What is available today?
Signup and a free sample are offered through Cupel. Current enrichment outputs and their limits are listed above. Some signals use sampled frames; a completed sample does not mean every source frame was processed.
Does this produce robot commands?
Do not treat preview estimates as robot commands. The example journey describes product direction, including target mapping and policy evaluation. It does not establish that an end-to-end retargeting or learning workflow is available or verified for your robot.
What if I already have a robot dataset?
Contact us with the format, size, available observations and actions, and the checks or export you need. Dataset repair, release decisions, and policy-utility comparisons need evidence for your source and target. They are not promised by the preview enrichment offer.
How should I assess an output?
Inspect the actual files and frame coverage. Distinguish inferred labels from measured geometry and review errors before training. Whether a dataset change helps a policy requires a separate comparison against a baseline under matching conditions.
What does your
dataset need?
Tell us what data you have, which robot or training setup you are targeting, and what output you need.
Discuss your dataset →