Useful robot training data starts with a question: what should a system learn from this example? A phone recording of someone serving water can show object identity, action order and a visible outcome. It does not automatically contain the measurements needed to control a robot hand.

Make local context observable

Consider the difference between placing a sealed water sachet and placing a rigid bottle. Both might receive the label “pick and place”, but the object behaves differently. A useful annotation identifies the sachet, the tray, the handling constraint and the result: two intact sachets on the tray, without a visible leak.

Language adds another layer. Preserve the instruction a contributor actually used, including Pidgin where appropriate. Write down the intended meaning and relevant scene context separately. An English rendering should help another reviewer understand the task without erasing how it was originally expressed.

Start with a small, documented collection

  • Choose one repeatable task and define an observable success condition.
  • Record permissioned examples with the relevant objects and action visible.
  • Annotate timed steps, object names, instructions and outcomes.
  • Keep contributor references, recording permission and reuse terms with the annotations.
  • Ask another reviewer to inspect both the clip and its labels before treating them as reliable evidence.

Include variation deliberately: object placement, phrasing or lighting might change while the task stays the same. Record unsuccessful attempts too. A collection containing only polished examples can conceal the situations a system will encounter when an object is missing or an instruction is ambiguous.

Define the contribution precisely

Nigeria-specific examples can improve coverage of particular objects, language and settings. That does not mean African activity is absent from existing research. Ego4D documents a global egocentric-video effort and includes CMU Africa among its participating institutions. Its attention to metadata and privacy is relevant when designing a smaller collection.

In the 9jaBots task studio, a local video can become timed annotations with context and provenance. Templates are authored starting points, not recorded evidence. The exported JSON references the source video; it does not include the video or robot action trajectories.

For initial evaluation, inspect the authored instruction sample and its dataset notes. Keep that sample separate from any future human-recorded collection. For physical control research, plan the additional robot observations, action measurements and calibration that the intended learning method requires.

Bring people into the project

Use 9jaTesters to request collection, annotation, testing or human review. Describe the task, volume, location or languages, and acceptance criteria in your brief. For contributor opportunities, join the 9jaTesters workforce.