Overfit the World.
Real-world data for physical intelligence
We help world-model labs, robotics, autonomy, and embodied AI teams build the data systems, evaluation benchmarks, and operational playbooks required to make models work in the real world.
Overfitting is a bug goal.
When the dataset is the world, overfitting isn't failure — it's understanding. We have are building the whole world, one captured corner of reality at a time.
By the numbers
- / The frontier01
10Mhours
Overfit the world.
Ten million hours of real-world experience: the data a model needs to stop guessing at physics and start knowing it.
TARGET / REAL-WORLD EXPERIENCE
- / Accumulated experience02
1Mhours
A foundation, not a finish line.
- / Monthly capacity03
500Khours / month
Research-grade, at industrial pace.
- / Scene diversity04
10Kscene types
Diversity built in, not filtered in.
- / Experience density05
9signals per hour
Beyond raw video.
Latest research
TouchScale: 500 Hours of Human Vision and Touch
TouchScale is a large-scale RGB-tactile dataset for embodied AI and robot learning, containing about 500 hours of synchronized human visual and tactile data. It captures first-person and wrist-view video together with hand contact and pressure, supporting contact prediction, action recognition, and robot manipulation training.
Read →What We Do
Capabilities
Build With Us.
If you are building models that need to understand, evaluate, or act in the physical world, we would like to talk.