Client B · Physical AI & robotics
2,000 robot videos described every week by a team built in seven days
Natural language descriptions of dual arm robot video for foundation model training, inside the client’s own labelling interface.
Illustrative image, not client data
The challenge
Client B builds foundation models that give robots general purpose physical skills. Its models needed precise, consistent text descriptions of household tasks such as bussing tables and folding laundry, captured on video.
Each video carried two camera streams, one per robot arm, so annotators had to follow concurrent actions across both feeds. The volume was about 2,000 videos a week, and each video had to be finished within five times its own length.
What we did
We assembled and trained a team of more than 30 annotators within one week, running three shifts to hold the weekly throughput.
Instruction documents were written with the client so every annotator used the same vocabulary, and we fed back on the client’s custom interface to make it faster to work in.
Results
About 2,000 videos processed every week, with every throughput and time based KPI met.
The engagement closed when the client changed its internal approach, with an open invitation to return.
