The Physical AI Data Foundry
Not by hand.
But automated.
Today, most robotics data is collected through teleoperation — a human operates a robot, one task at a time. It's slow, expensive, and impossible to scale. And none of it happens in kitchens.
The Lili-o foundry runs household tasks autonomously across 50+ purpose-built home environments — kitchens, living rooms, bathrooms, laundry rooms — robots operate 24/7, retry on failure, and generate episodes continuously. No human operator. No wage floor. No ceiling on scale.
The market settled for trade-offs. We didn't.
| Simulation | Human-Centric | Téléopération | Lili-o | |
|---|---|---|---|---|
| Rich Metadata | Low | Medium | High | High |
| Environment Diversity | High | High | Low | High |
| Price | Medium | Low | High | High |
| Cross-embodiment | No | Yes | No | Yes |
| Scalable | High | Medium | Low | High |
| Companies | Lightwheel · NVIDIA | Scale · Senseirobotic | Tutor · Figure · Agibot | Lili-o |
*EU AI Act compliant
One use case. Deployed everywhere. Running forever.
Each use case is a self-contained, hardware-agnostic program that runs on any compatible robot without modification. Build it once — deploy it across the entire fleet.
Hardware agnostic
One use case, any robot. Unitree, Rainbow, Agibot — no reprogramming.
One-Shot learning
5-minute demonstration → autonomous skill. Industry standard: 1 hour+.
Learn how the One-Shot method works →Perpetual output
One build. Infinite episodes. 24/7 generation across the entire fleet.
The modalities that actually matter.
Force-torque and proprioceptive signals are absent from almost all existing public datasets. This is the gap that separates models that look like they're manipulating from models that can actually feel it.
RGB Video
Multi-view synchronized capture
Depth Maps
Aligned to RGB at every frame
Tactile / Force-Torque
The signal almost no dataset has
Joint Encoders
Full proprioceptive state
Episode Metadata
Auto-labeled, structured, ready-to-train
5 modalities. Synchronized. Auto-labeled. Cross-embodiment compatible. EU AI Act compliant.
Real homes. Real people. Real tasks.
Our second collection channel sends instrumented participants into their own homes — kitchens, bathrooms, laundry rooms — wearing RGB-D cameras and haptic gloves, performing everyday household tasks as they naturally would.
This captures the environmental chaos, behavioral variance, and physical interaction that a controlled environment can never replicate. The mess on the counter. The wet dish. The awkward cabinet angle.
- Dish washing
- Meal prep
- Appliance use
- Counter cleaning
- Object sorting
- Table setting
- Tidying
- Vacuuming
- Folding clothes
- Loading washer
- Ironing
- Sorting laundry
- Surface wiping
- Bin handling
- Towel folding
- Cleaning fixtures