What the work actually involves

You watch short clips — a robot arm approaching a fixture, a human hand seating a bearing, a gripper mis-closing on a fastener — and label what is happening with engineering precision. That means naming components correctly (a shoulder bolt is not a cap screw), marking where contact begins and ends, distinguishing a rotation about the wrist axis from a translation of the whole arm, and segmenting a task into the sequence a process engineer would recognize. A second, equally large part of the job is review: reading other annotators' work and flagging where a label is wrong, inconsistent with the guideline, or plausible-looking but physically impossible.

The value you add is not speed. It is that you know a compliant gripper deforms under load, that a part settling into a chamfer is a distinct event from insertion completing, and that occlusion in frame 40 does not mean the part vanished. Robotics teams train on this data; systematic mislabeling of contact events or motion types propagates directly into model behavior.

What the screen looks for

  • Real mechanical depth under follow-up — kinematics, degrees of freedom, fits and tolerances, common fasteners, tooling, assembly sequence.
  • Whether you can describe a physical interaction in unambiguous words rather than gesture-level vagueness.
  • Judgment about ambiguity: what you do when the guideline does not cover the clip, and whether you escalate with a written example or just guess.
  • Consistency discipline — applying a rule the same way on clip 300 as on clip 3, even when you privately disagree with it.
  • Plain, careful written English. Annotation notes and disagreement rationales are read by people who were not in the room.

Logistics

Fully remote and contractor-based, with asynchronous task queues rather than fixed shifts. Expect an onboarding guideline document, a calibration set scored against gold answers, and periodic guideline revisions that require re-reading rather than relying on memory. Volume tends to arrive in project waves; candidates who can commit a predictable weekly block get steadier assignment. A reliable machine and connection matter — video review at frame-level scrub is bandwidth-sensitive.