
Deployment & research for Physical AI. Backed by @ycombinator.
San Francisco
Joined March 2024
- Tweets6.3K
- Following16
- Followers227K
- Likes10.4K
Pinned Tweet
Today we're open-sourcing HubEgo-100K: the most diverse egocentric dataset on 🤗 @huggingface.
100,000 hours of skilled tasks from 100,000+ different Hub contributors.
Free for research and commercial use 🧵
Hub is committed to accelerating research in the field by becoming the largest open-source provider of Physical AI data.
New datasets will keep rolling out on a regular basis: from advanced human demonstrations, to teleop data, to on-policy rollouts in deployment.
We're starting HubEgo-100K with a first release (v1) of 10,000 hours. The remaining 90,000 hours will roll out in batches over the coming weeks.
We can't wait to see what the community is gonna build with it.
Access it now: 🤗
huggingface.co/datasets/HubL…
good: we trending on X
groot: X confuses us with @NVIDIARobotics 🥸
Robot policies work in the lab. The real world is where they break.
To close that gap, we are open-sourcing the most diverse egocentric dataset with the help of @LeRobotHF.
Coming to 🤗 @huggingface on Thursday November 12th.
We let our contributors describe Hub instead 🎥
No script, just people in Brazil talking about recording their everyday life.
The detail that matters is the habit. They record their daily tasks every day, because it fits a life they already live.
Supply is the hard side of any physical data network.
Incentive-driven collection churns when the incentive ends. A habit compounds into collection capacity that keeps growing.
That retention lets Hub keep adding new environments and tasks without starting from zero.
Your job pays you once. Now it can pay you twice.
Commercial tasks are live on Hub. Until today, you earned by recording tasks at home.
Now you can record your actual job: the shop, the salon, the farm, the workshop, the factory and more.
Same hours you already work. The difference is they count now, at a higher rate than household tasks.
What do you do at work that a robot could never do?
🧺 A robot can be taught to fold one towel. It cannot be taught to fold laundry.
No two piles land the same way. There is no motion to memorize.
That gap is the whole game in robotics, and it only closes with footage of real hands working in real homes.
If you record with us, you are already closing it.
So tell us: what has Hub taught you about robotics and where AI is heading?
🇧🇷 One of our contributors just aired on Brazil's biggest television network.
A news segment. No ad spend. Her, recording her own day with Hub, getting paid for it.
For the first time in history, anyone with a phone can help teach machines how the physical world works.
Thousands of people already are.
This is what it looks like before everyone else catches on.
Hub creators have now recorded over 2,000,000 clips of everyday human tasks.
At home: folding laundry, weeding a garden, taking care of animals.
On the job: mechanics, agriculture, construction, and many more. This side is scaling fast, and will soon cover any kind of work a human can do.
Real people, in the wild. First person, across the Americas.
This is the physical world getting written down for the first time, and it is the one thing robots cannot learn from the internet.
What could you teach a robot that nobody else can?
AI has seen a dog get fed more times than anyone could count. It still couldn't feed one itself.
Recognizing a task and knowing how to physically do it are not the same kind of intelligence. Most AI training data teaches the first. Almost none of it teaches the second.
What that second kind actually requires is specific: the half-second a hand waits before releasing food, timed to the exact moment a dog's mouth is ready, not a second before.
The angle of a crouch that keeps balance on uneven tile.
The grip that's firm enough to hold, loose enough to let go cleanly.
It only exists in the motion itself.
Hub pays real people to record exactly this, first person, inside their own homes, at a scale no single lab could stage on its own.
One woman, one dog, one ordinary Tuesday, is a single recording. Multiply it by everyone else doing the same ordinary things in their own homes, and it stops being a data point.
It becomes the dataset embodied AI has been missing.
Founding Engineer, France.
The systems this person builds ship data to frontier AI labs.
Remote, open now ⬇️
This quoted post is unavailable.
🇧🇷 Hub is hiring a Field Operator in São Paulo
Full ownership of the ground there: closing businesses, installing kits, training crews, keeping the recordings flowing.
We hire on drive, not resume, this can be someone's first real job.
Know someone built for this? Refer them and share this post. If we hire your referral, you get $500.
Apply on our profile.








