Pegasus 1.6 by TwelveLabs provides video intelligence for physical AI, turning first-person and robotics footage into structured, training-ready data through action segmentation, dense captioning, quality scoring, and compliance flagging.
What the product does and how it is positioned
Pegasus 1.6 by TwelveLabs is a video intelligence solution engineered to convert first-person and robotic video into structured, training-ready data for physical AI pipelines.
The platform enables teams to segment actions, generate dense descriptions of scene context, score visual quality, search libraries using natural language, and flag sensitive personal data before review or downstream processing.
Source-supported ways to use the product
Robotics laboratories can transform raw first-person and robot footage into structured training datasets with timestamped actions and interaction captions.
Commercial data providers can standardize delivery schemas, score clip quality, and verify consent requirements prior to buyer delivery.
Physical AI teams can query extensive video archives using natural language to uncover long-tail anomalies, rare events, and duplicate scenes.
The documented workflow, where available
Identify valuable moments and filter raw footage against specific task requirements before sending it downstream.
Check footage against behavioral, visual, and privacy criteria, assessing stability, framing, occlusion, and PII presence.
Return structured, timestamped JSON outputs that teams can review, filter, and ingest directly into machine learning workflows.
The platform converts raw egocentric and robot video into structured data aligned with custom task structures. It identifies discrete, timestamped steps that can be mapped to user-provided verb lists and action hierarchies.
For language-conditioned robotic policies, the system generates natural-language descriptions that detail hand-object interactions, spatial relationships, and contextual scene dynamics without assembling separate transcription or framing tools.
To keep low-quality recordings out of model pipelines, the solution evaluates footage stability, framing, visual occlusion, and action clarity using configurable quality prompts.
The screening process also incorporates automated compliance checks that locate minors and personally identifiable information, including faces, license plates, badges, and documents, prior to delivery.
Checks to run with your own material and workflow
What was checked and when
Answers based on the source-checked product record
The platform processes first-person egocentric video and robot recordings to prepare training data for physical AI workflows.
Yes, teams can bring their own verb lists, action hierarchies, and granularity requirements to map detected video actions.
The system scores footage based on clip stability, framing, visual occlusion, and action clarity before manual review.
The platform flags personally identifiable information such as faces, license plates, documents, and badges, as well as minors.
Outputs are delivered as structured, timestamped JSON records accessible through an official API and an SDK with async task handling.
