Who hires the people who train robots, read off their own job boards.

Listed companies in the role files today, each with its exchange and ticker in platforms.csv: Amazon Prime Air (79 roles), Avride (65 roles), Serve Robotics (22 roles), Aurora Innovation (18 roles), Symbotic (14 roles), DoorDash Labs (11 roles), Kodiak AI (11 roles), Waymo (11 roles), Appen (9 roles), Latitude AI (6 roles), Torc Robotics (5 roles), Zoox (3 roles), ALTEN Technology USA (2 roles), Flexdrive (Lyft) (1 role), Innodata (1 role), TELUS Digital AI Community (1 role) and Wing (1 role).

• Once a day we read each employer's own public job board: its applicant tracking system's feed, or the job list on its own site. The day each employer's first read happened is in the coverage table.
• A role is in the role files only when we read it from a job board that anyone can open with no account and no automated check to get past. An employer we can read only some other way stays on our public job board and in the market's daily totals as they were recorded, and never appears role by role. Tesla: Its careers site turns away every reader that is not a full web browser, so we read it with one. That is why the jobs up on our board, 1,196 today, are more than the role files hold up: the board also lists 171 at Tesla.
• A row is one role. An employer that posts one job once per city or reposts it at the same pay is kept as one role, with all of its places.
• Pay is the posting's own pay line, kept word for word. Hourly rates and ranges are also read into USD figures; salaries and pay per task stay as written.
• An employer's roles in its first week with us were already up when we started reading its board, so they are never counted as newly opened (the first_week column of the employer-by-day file).
• A role comes down on the first daily read that no longer lists it; last_seen_at is the last read that did.
• Until the change log began, a role kept only its latest state: its pay before a change, and the days a role that came back was away, were not kept. From then on the change log holds every one of them.
• When we change which postings of a board count as robot-training work (a new rule, or a board read in full instead of in part), roles can enter or leave our records on that day. Those days are not marked apart in the files yet, so a jump in one employer's new or closed roles on a single day is worth checking against its roles.
• Rebuilt from our role records, the count of jobs up matches the count recorded on the day within a median 0.2% over 48 recorded days. The widest gap, 1.1%, was on 13 Oct 2026. Where the two differ, it is roles counted that day that are no longer in our records, or roles taken down later whose board went unread that morning; the market-by-day file puts both counts side by side for every day.
• The kinds of work are our own classification, applied the same way to every employer: Robot Teleoperation, Data & Demonstration Collection, Egocentric & Wearable Capture, Embodied-AI Data Annotation, Robot Task Evaluation & QA, Robot Fleet Operations, AV & Drone Operators.