Jim Fan open-sources Flora, a robot video labeling engine
The models already label 200,000 hours of video a day and feed NVIDIA GR00T training.
Jim Fan has open-sourced Flora, a suite of models that reads raw video and turns it into training signal for robots. He says it hits Astra-level performance at 125x lower cost, and the weights are public as of today.
We created it to scale the frontier for robot learning. Now you own the weights too.
The system is not a research artifact sitting on a shelf. Fan says Flora already powers an in-house data engine labeling 200,000 hours of video a day on 500 GPUs. Yunze Man, who posted the technical thread, says the team uses it to train NVIDIA GR00T.
What it actually outputs
Three things, according to Fan. Timestamped captions, used to curate and filter human egocentric video (footage shot from a head-mounted camera) and robot data. Dense progress rewards, for quality control and RL post-training (reinforcement learning applied after the base model is trained). And rubric scoring for real-world benchmarking.
Man describes it more simply. The model watches robot and human video and annotates what happened and how well it went, including failure reasons.
That last part is the interesting bit. Labeling video with what a person did is one problem. Scoring how well it went, frame by frame, is what you need if you want to filter a giant pile of scraped footage down to the clips worth training on, or to hand a policy a reward signal without a human in the loop. Data curation is the bottleneck that nobody gets to skip, and it is usually done with a mess of internal tooling that never ships.
Shown versus claimed
The 125x cost number is the team's own framing, not an outside benchmark, and the sources do not break down how it was measured. The 200K hours per day figure is also self-reported. What is verifiable is the part that matters most here. The weights are out, so anyone with GPUs can run Flora on their own video and see what the captions and rewards look like.
Fan pointed to a website, an open model suite and a technical thread from Man for the details.
It watches robot and human video and annotate what happened (timestamped instructions) and how well it went (rewards and failure reasons).
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.