The brief
- JEPA models predict world evolution from robot actions, learning physics from video without text.
- Closed $1.03B seed round at $3.5B valuation in March 2026, backed by Bezos, NVIDIA, Samsung.
- Targets robot learning, warehouse automation, and industrial reasoning. Offices in four continents.
Technical approach
JEPA
Joint Embedding Predictive Architecture
Models predict future video frames from robot actions without generating text, producing compact representations for planning and control.
Self-supervised
Learning from interaction and video
Models learn from internet video and robot trajectories. Generalizes to new environments with minimal task-specific training.
Applications
From robotics to industrial automation
Same model deploys across robot types and industrial tasks: manufacturing, aerospace, and biomedical.
