Helping robots learn from the real-world Founder- https://xitter.cf/t.co/EyONynBAXw prev- @CUHKofficial
sz - sf - blr
Joined October 2011
- Tweets2.2K
- Following4.6K
- Followers738
- Likes2.3K
Pinned Tweet
Human skill is the pre-training data for robots.
@theopenrobot builds the hardware layer for this collection infrastructure.
The Last Apprenticeship explores what happens when that experience scales from millions of hours to billions.
Many of these ex-people love what we cook.
a lot of companies are getting funded. All of them are incredibly technical, ex-Google, ex-OpenAI, Berkeley, Stanford, etc. all of them are techno-optimists, which I love.
at OpenRobot we're different: we don't fancy degrees. Most of us don't even have a degree and have never worked for anyone else, let alone Google.
we want to stay non-conformists, we operate differently compared to current companies/startups
will share more about it someday
Don't sleep on EGO.
there have been like 20 times over the past year people saying slop about it.
but it keeps winning. LOL.
Human skill is the pre-training data for robots.
@theopenrobot builds the hardware layer for this collection infrastructure.
The Last Apprenticeship explores what happens when that experience scales from millions of hours to billions.
EPIC 💦
We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing Flora, a suite of open models to parse the physical world and turn raw footage into rich training signals.
Flora already powers our in-house data engine, labeling 200K hours of videos a day on 500 GPUs. We created it to scale the frontier for robot learning. Now you own the weights too.
Flora gives you:
- High-quality, timestamped captions to curate and filter human egocentric and robot data;
- Dense progress rewards for quality control and RL post-training;
- Rubrics scoring for real-world benchmarking
Try it today, links in thread:
who has got the most cost-effective quality UMI?
would like to buy 1.
@almond_robotics how much is your UMI?
do u ship to india?
who has the best stereo-device-game in the world?
ranked on-
- useful data
- operator seamlessness
- cost
Over 3M hours of egocentric data now exist:
Build AI: Egocentric-1M
Midcentury: 2M+
Lightwheel: 100K open
Impressive scale. But look at what's inside. Lightwheel's own task mix: 62% is cleaning, tidying, packing and inventory. By my count, under 18% is genuinely hard, skilled-hands work.
Frontier models already one-shot easy and medium tasks, and simulation is far ahead in this category.
My bet: ~10% of the ego data collected so far is actually useful henceforth for the next gen of models.
So ego isn't saturated. It's starting from scratch. The only data that matters now is hard: precise tool control, judgment, fitting to a line, controlling a process.
And the price of that data? It goes very high.
