The future needs data from everyday people.
Training-ready multimodal human-activity datasets for physical AI. Built by international and national gold medalists in math and robotics.
Every physical AI company needs a foundation model first.
Before any task-specific training, robots learn broad manipulation, navigation, hand-object interaction and spatial understanding. That's 90–99% of their training data. It has to come from real people.
Founded by NTU robotics researchers.
RealAI was started by researchers from Nanyang Technological University in Singapore, and works with NTU's MARS Lab on physical AI.
ShanghaiRanking
QS World Rankings
NTU MARS Lab
NVIDIA-backed research in physical AI, multimodal LLMs and robotics.
- NVIDIA Academic Grant for multimodal robot manipulation, 2026
- Papers at ICLR, CVPR, ICRA and RSS
- ICLR oral, top 1.8%
- Meta Project Aria participant
- ECCV challenge winner
MARS Lab collaborators







Harvard
UC Berkeley
Stanford
Oxford
Princeton
No saturated tasks.
For the first time ever, capture isn't happening in staged rooms and warehouses, with the same tasks repeated until the hours pile up. Models can now learn in homes people actually live in.
Hours stay balanced across every task.
Multimodal data capture.
We record video, hand motion and 3D room data at the same time, from the wearer's point of view. All streams are synced.
Request sample dataVisual
The task from eye level, exactly as it happened.
Motion
Every finger joint, grip and reach, tracked through the whole task. The fine dexterity robots need to learn.
Spatial
Room geometry and object positions, mapped as the person moves.
We own the entire pipeline.
The hardware, the app and the data engine are all built and run in-house.
-
01 / WearCustom hardwareBuilt in-house.
Contributors wear our hardware and go about their day at home.
-
02 / RecordRealAI appRecord everyday tasks.
The app assigns tasks, so every gap in the dataset gets filled on purpose.
-
RawProcessed03 / ProcessData engine
Turn it into data.
Every session is synced, privacy-blurred and labeled.
-
04 / TrainFoundation models
Train physical AI.
Models learn from real people in real homes.
Quality at scale.
Our hardware is built for mass production, so hours grow with demand and quality holds.
Usable data rate.
From idea to production and a sample dataset.
