TwelveLabs aims Pegasus 1.6 at first-person video to speed up robot training data

TwelveLabs said in a press release on Oct 6 that Pegasus 1.6 is its first video model trained to understand egocentric, first-person footage, a category earlier versions handled poorly. The model supports workflows including timestamped action segmentation and labeling, dense captioning and quality scoring, aimed at turning recordings of human work into structured datasets for robotics and other physical-AI developers; the company cited an estimate that one hour of raw video takes about 155 hours of manual labeling.

Customers and pricing for the robotics use were not named. The launch targets the same human-demonstration data supply that Mecka is building with the $60 million Series B MEKAWIRE covered on Oct 7.