What it is
Mission-based capture app
How it works
Contributors record predefined task scenarios
Rewards
On-chain, per accepted episode
Status
Live
The internet of the
physical world.
Most robot training data is footage machines can only watch. ZenO captures first-person video with metric 3D coordinates. Data robots can actually move with.
Scroll
01 / Problem
Robots are 10,000× behind.
LLMs train on trillions of tokens. The largest robot datasets barely cross a billion frames. Most of it is unusable: web-scraped video with no 3D coordinates, lab data that never leaves the lab, academic sets locked behind non-commercial licenses.
The problem isn't collecting more video. It's collecting video a robot can learn motion from, legally.
Scale Gap
1.5T+vs~1B
LLM training tokens vs. frames in the largest robot datasetsLicensing
Most
public hand datasets are non-commercial onlyCost
$200+/hr
to run traditional teleoperation rigs02 / Solution
Three channels, one standard.
ZenO runs a global contributor network across three capture channels, each producing a different layer of what robots need.
01ARKit Capture
iPhone LiDAR records metric-scale 6DoF camera pose
Per-frame depth, synced to every frame
No scale ambiguity, no guesswork
02Multi-Cam Rigs
Head-mounted 360° plus wrist cameras
Run by trained contributors
End-effector trajectories in real homes & kitchens
03Sim Teleoperation
Drive a virtual robot arm with your phone gyroscope
Joint states logged at 30Hz
Directly trainable action data
iPhone LiDAR records metric-scale 6DoF camera pose
Per-frame depth, synced to every frame
No scale ambiguity, no guesswork
Everything lands in one format: LeRobot, with per-channel statistics, ready for training.
03 / Why ZenO
What others can't ship.
Coordinates, not just pixels.
Crowdsourced video without metric scale is just footage. ARKit LiDAR gives every frame an absolute position in space.
A license you can actually use.
Every contributor signs explicit consent at capture. No scraping and no academic-only limits. Commercially clean from capture to delivery.
Scenario-driven, not random.
Every recording follows a predefined task scenario with multiple takes. Structured, comparable, labeled.
We eat our own data.
ZenO Lab trains virtual robot arms on our own datasets. If the data doesn't move a robot, it doesn't ship.
04 / Traction
Already in the wild.
Demonstrations
12,763
teleop demonstrations · joint states at 30HzEgocentric Video
820+hrs/mo
captured every month across 23 task domainsOperators
145
contributors across the globePublic Catalog
Live
on Hugging Face · free sample tier05 / Products
Three products, one pipeline.
Capture, processing and proof: the three layers that turn contributor recordings into training-ready datasets.
Try before you license.
Sample episodes are free on Hugging Face. For full datasets and custom collection specs, talk to us.