Nvidia open-sources Flora, cutting robot video labeling costs by 125x

Nvidia says its new open-source labeling engine Flora annotates an hour of robot video for about 72 cents — compared to roughly $88.50 with OpenAI's GPT-6 Astra. The company released the system openly and says it already annotates up to 200,000 hours of video per day across 512 GB300 GPUs during training of its GR00T robot models. The models are available on Hugging Face.

The price gap is the headline, but the mechanics matter more. Robotics models learn from video, and that video needs text labels — lots of them. Flora's real pitch is that it produces all three kinds of labels the training process needs, at a cost that makes million-hour datasets feasible.

Why does robot video need three kinds of labels?

According to Nvidia's GEAR research lab, large AI models for robots require three types of annotations. Most existing tools provide only one, which forces labs to stitch together incompatible pipelines.

The three label types:

• Action descriptions: what happens when in the video, used to select training data.
• Progress ratings: how well a robot advances step by step, so it can learn through trial and error.
• Success judgments: whether a task was completed, used to monitor robots during deployment.

Large language models like GPT-6 Astra can generate several of these label types, but Nvidia argues they're too expensive for the massive volumes of video involved. That cost problem is what Flora targets.

What is Flora actually built from?

Flora consists of three fine-tuned Qwen3.5 models, each with a specific job. The design keeps the pipeline modular and cheap to run.

• A 9B model splits videos into action segments.
• A 27B model describes those segments, factoring in what happened before and after each action.
• A second 27B model, called Flora-Rewarder, takes on three roles depending on the prompt: it rates an episode's success on a scale of 1 to 5, classifies each half-second step as normal, suboptimal, or failure, and assigns scores based on self-defined rubrics.

That third model is where much of Flora's value sits. Judging whether a robot failed midway through a task — not just at the end — is something almost no competing model does well.

How big is the cost advantage really?

In a test with 50 videos, Flora scored slightly better than GPT-6 Astra, according to Nvidia, even though OpenAI's model was allowed to use its full reasoning capacity. Claude Opus 5 scored much worse in the same test and still cost about $42 per hour of video.

The comparison numbers:

• Flora: about $0.72 per hour of video.
• GPT-6 Astra: roughly $88.50 per hour.
• Claude Opus 5: about $42 per hour, with weaker results.

There's a caveat worth flagging. The Flora cost figures are projections, not production measurements. Nvidia calculated them based on an assumed rental price of $10 per GPU per hour. The 125x figure is the company's own estimate, and readers should treat it as such.

Where does Flora win and where does it lose?

The benchmark picture is mixed, and Nvidia's own paper acknowledges the weak spots. On the public WGO-Bench, Flora is best at detecting action boundaries but trails Astra in the overall score and falls behind Gemini 3.8 Flash in pure description accuracy. The researchers attribute Flora's gains mainly to temporal localization — knowing exactly when an action starts and stops.

On Nvidia's own DenseCaptionBench, Flora leads across all metrics for human recordings. For robot videos, it comes in second behind Astra. There's also a methodological catch the researchers acknowledge: GPT-6 Astra evaluates the descriptions while also being one of the models under test.

The reward results are more one-sided. On RoboRewardBench, Flora's error rate is 27.8 percent lower than Astra's. The paper attributes the gap to partially completed tasks: when a robot only tips a bottle over instead of knocking it off the table, Flora assigns 3 out of 5 points to match the human reference, while Astra gives just one.

Flora-Rewarder catches 53.3 percent of failed steps within an episode. No competitor gets above 11.7 percent, and Astra doesn't catch a single one. The tradeoff: Flora falsely flags about 11 percent of normal steps as errors. Astra still leads on rubric-based scoring and trajectory ranking.

What does it look like in practice?

In a video with many short hand movements, Flora identifies 34 of 37 individual actions, while Astra catches only four. That's the difference between a usable training label and a blur.

But Flora has visible failure modes. It sometimes splits longer actions into too many small steps, and it occasionally gets descriptions outright wrong. In one video, a robot hand wipes a plate with a cloth, and Flora describes it as folding a napkin. Anyone building on the system will need to budget for that error rate.

Why is Nvidia giving this away?

Nvidia positions Flora as an open alternative to commercial services like Scale AI's Physical AI Data Engine. Researchers can use it to scale their own robotics datasets and test approaches like transferring human demonstrations to robots.

The giveaway fits a pattern. Nvidia is investing billions in open AI models to keep developers in its hardware ecosystem and drive demand for its chips, and Flora serves that same goal. The 512-GPU production deployment makes the point plainly: labeling at this scale requires Nvidia silicon.

The company has also spent years trying to offset the shortage of robotics training data through simulation and automated data processing. At GTC earlier this year, Nvidia showed a data pipeline that automatically prepares, multiplies, and scores training material. Flora adds detailed labels and reward signals from video to that pipeline.

What does this mean for the robotics market?

All of this feeds Nvidia's strategy of covering the entire robotics value chain — hardware, training data, simulation, and the AI models themselves. Flora closes a gap in the data layer that competitors were selling access to.

The competition is already moving. Google's recently announced Gemini 4 Argon targets a similar part of the robotics training process, also identifying work steps in training videos and flagging failed actions. With open tooling on one side and frontier models on the other, the cost of labeled robot video is falling fast — and whoever labels the world's robot training data will shape what those robots learn.