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:show more

Flora website: Open model suite:

Check out Yunze's technical deep dive:

whoa!

Open model drop!

Nice progress

Incredible work team, excited to try it out

Please do!

Congrats Jim!

❤️❤️

oh goddamn this is insane!!

Dense progress rewards are the part I'd watch. Captions tell you what happened, but a per frame score of how close the task is to done is what lets you throw out the bad demos before they poison a policy.

Open weights change the calculus: the bottleneck moves from labeling to footage collection. With 200K hours a day on 500 GPUs, robotics is finally learning at a scale that echoes the internet-era data playbook — and doing it at 125x lower cost is the real story.

200K hours a day on 500 GPUs is 400 hours of footage per GPU per day, about 17 times real time. At that throughput the bottleneck moves: it is no longer collecting video, it is which clips the captions and progress rewards decide to keep.

Incredible wanted to know which specific robotic arms you use for this demonstration?

A 125x cheaper data engine is a strong claim; the real test is whether its dense rewards work on robots it never saw.

125x cheaper labeling means the robots finally get to watch more youtube than we do

Congrats @DrJimFan Are these benchmarks / evals open / public ?

Should be!

Flora's raw-footage-to-training-signal pipeline is exactly the data layer Genesis Chassis has been starving for. Open weights at Astra-level fidelity changes the embodied stack calculus overnight. ⚡

Robotics is short on labeled data, not robots. If Flora really turns raw footage into training data at 125x lower cost, the edge moves to whoever owns the most video. Who's sitting on the biggest pile?

every shaky first-person cooking video out there just got promoted to robot homework

I’d love to see what this makes learnable from the awkward bits: the missed grasp, the pause, the second attempt. There’s a lot of practical knowledge in how people recover.

The dense rewards are the part I'd watch closely. Can Flora distinguish genuine task progress from motion that merely looks productive? A comparison of downstream robot success with and without those rewards would make the labeling gains much more tangible.

Do the progress rewards hold up when the task needs backtracking, like undoing a bad grasp? Most progress scores I've seen just tank there

the performance to cost ratio is crazyy 🫡. congrats!!

200K hours of video a day on 500 GPUs is about 17 hours of footage per GPU every hour. At that rate, is the bottleneck now video decode or the model itself?

Labeling vendors who charge by the hour for robot footage are about to have an awkward budget meeting.

200k a day crikey

How noisy do the dense rewards get during heavy occlusion?

Very diplomatic of Astra to say the toy plane's wing fell onto the table and still call it suboptimal. Explains the 0.0 failure recall

200k hours a day on 500 gpus is ~17x realtime per gpu for dense timestamped captions. what fps are you sampling at?

Flora's real flex is the throughput line. 200K hours of video a day on 500 GPUs = 400 hours per GPU per day. Each GPU labels ~17 hours of footage for every hour it runs. At that speed, the bottleneck for robot learning moves upstream, to collecting the footage.

Open-sourcing the weights is a big deal for robot learning.

$0.72 an hour for open weights is generous

200k hours a day on 500 GPUs is a real number. curious what the labeling error rate looks like on messy egocentric footage

NVIDIA to the moon man wtf

Open-sourcing models that turn raw footage into useful training signals could be a big step for robotics. Better data quality and lower costs are exactly what can help accelerate real-world robot learning.

Can Flora's progress rewards capture setbacks too, like a robot undoing a step it just completed?

125x cheaper is insane!