TwelveLabs releases Pegasus 1.6 with egocentric video support for robotics training

TwelveLabs launched Pegasus 1.6, a video understanding model with specialized support for first-person (egocentric) footage, targeting robotics developers and data suppliers. The model converts recordings of physical work into labeled, timestamped training data.

Building on Pegasus 1.5's video description and segmentation capabilities, the new release adds training specifically for first-person video. The company said earlier versions handled this category poorly. Pegasus 1.6 is designed for native temporal reasoning and end-to-end video understanding, identifying the order of actions, objects involved, which hand performs the action, and when actions start and end. It turns raw video into structured time-based metadata.

The release also adds native image analysis and improved person and object identification, allowing customers to process still images through the same API as video. TwelveLabs outlined five primary workflows: breaking recordings into labeled actions, generating detailed captions, screening footage for quality, finding unusual events and duplicates, and flagging potentially sensitive material before downstream use.

Pricing remains unchanged from the predecessor: $1.75 per hour of input video, $3 per million image-input tokens, and $7.50 per million output tokens. Enterprise contracts use custom pricing.

Co-founder and CEO Jae Lee said the company previously focused on media, entertainment, sports and advertising, and the move into robotics applies its capabilities to physical behavior recognition. Noting that robot models are "super data hungry," Lee said first-person recordings can serve as a demonstration source, but Pegasus provides descriptions and rough trajectory understanding; pressure, touch and precise control remain separate problems that require combining with other data and control systems.