Research

Building models that
understand the physical world.

We are using the data we collect to train a world model and a foundation model for physical AI — building the research credibility to partner with frontier robotics labs. Research will be released soon.

Research preprint releasing soon

Research directions

I

WORLD MODEL

Learning the structure of the physical world.

IN PROGRESS

We are training a world model on real, egocentric human demonstration data — learning the geometry, object dynamics, contact physics, and cause-effect structure of everyday manipulation environments. Unlike simulation-derived world models, ours is grounded in real sensor streams: synchronized RGB, depth, IMU, and full-body kinematics captured across 25+ live environments.

II

FOUNDATION MODEL

A base model built for embodied agents.

IN PROGRESS

We are training a multimodal foundation model on top of our collected demonstrations. The model learns a shared representation across vision, depth, motion, and action — purpose-built for downstream robotics policy learning, imitation learning, and zero-shot transfer to novel tasks. Pretrained on DatraAI data; fine-tunable on your robot.

III

DATA FLYWHEEL

Better data, better models, repeat.

ONGOING

Model feedback continuously improves our collection protocols. As we train, we identify distribution gaps, edge cases, and underrepresented actions — then we go collect them. This closed loop between model evaluation and field collection is the compounding advantage that makes DatraAI's dataset increasingly more valuable over time.

Why we publish
research.

01

Credibility through publication

Research is the trust layer. Publishing rigorous work on world models and foundation models trained on our data demonstrates our collection quality and processing rigour to the labs we want to partner with.

02

Partnership with frontier labs

Our research output is the opening to co-development agreements, data licensing deals, and joint training partnerships with frontier robotics labs building the next generation of general-purpose robots.

03

A benchmark for physical AI data

By publishing models trained on DatraAI data, we create a public benchmark for what high-quality, hardware-synchronized, real-world demonstration data can achieve — setting the bar for the industry.

“The path to general-purpose robots runs through general-purpose data — real, embodied, and collected at scale.”

DatraAI Research