For AI & robotics teams

Define the observation.
Validate the signal.

Commission task-relevant human demonstrations for a learning or evaluation pipeline you control. Start with the requirements that determine whether the recording will be useful.

Before capture

What needs to be
in your specification?

Clear requirements prevent an attractive video from becoming an unusable dataset.

RequirementQuestions to resolve
Task & objectiveWhich action sequence, object state changes and outcomes matter? Training, evaluation or task discovery?
Coverage & variationWhich people, object variants, layouts, PPE, lighting and approved recovery cases are required?
Embodiment & observabilityWhat can human video support in your pipeline? Which signals require sensors, robot demonstrations or additional measurement?
Labels & synchronizationWhat taxonomy, temporal resolution, stream alignment and validation method will you accept?
Acceptance & deliveryHow are accepted hours counted? What quality thresholds, data schema and licensing scope are required?
06 / What the buyer receives

More useful than
a folder of footage.

Agree the schema and evidence needed to evaluate a batch. The deliverable is defined in the proposal, including what is not captured or measured.

Video, context and a review trail.

Accepted recordings can be paired with task IDs, recording conditions, camera information and review outcomes. Optional annotation layers are scoped against your requirements.

Human RGB video does not automatically include 3D pose, depth, force or executable robot trajectories. Sensor streams and derived labels require a suitable rig and separate validation.

Task boundariesObject labelsOutcome & retry tagsAgreed file schema

Technical scoping for AI teams ↗

collection-example/  # illustrative
  footage/           # accepted video files
  metadata/          # clip and capture records
  annotations/       # if commissioned
  review/            # QA decisions and reasons
  documentation/     # datasheet and agreed rights
  manifest.json      # agreed file index

A pilot is
an evaluation step.

Use representative footage to test your loader, assess visibility and inspect labels against real frames. Evaluate whether the dataset can support your intended use before expanding the collection.

No model-performance improvement, sim-to-real transfer or robot deployment outcome is guaranteed by recording human activity.

Scope a pilot

Illustrative pilot specification

Example scope for a controlled assembly task. Capture setup and outputs are agreed per brief.

Task
Align a part, insert it into a fixture and verify its position.
Variations
Part type, glove type and lighting. Agree the number of accepted episodes per variation.
Capture
Fixed close view as a starting point. Agree framing, frame rate, file format and synchronization before recording.
Acceptance
One complete episode shows the required steps and hand–object interaction, meets the agreed technical checks and excludes restricted content.
Delivery
Accepted files, agreed metadata, QA outcome and a file manifest. Annotation and additional sensor data require separate scoping.