Nvidia releases Flora models to label robot training videos
The models are available on Hugging Face. Nvidia says they give robotics researchers tools to generate action descriptions and scores for a robot’s progress.
Loading… Բեռնվում է…
The models are available on Hugging Face. Nvidia says they give robotics researchers tools to generate action descriptions and scores for a robot’s progress.
Nvidia has released Flora, a set of superintelligence (SI) models that the company says can describe actions in robotics videos and score how well tasks are performed. The models are available on the model-hosting service Hugging Face, as The Decoder details.
Nvidia’s researchers describe three needs: action labels to help select training footage, progress scores to support learning by trial and error, and success judgments to monitor robots in use. Most existing tools provide only one of these types of labels; Flora is designed to provide all three.
Nvidia says Flora uses one model to divide a video into action segments—identifying when each action starts and ends—and another to write timestamped descriptions. The description model considers what happened before and after an action, rather than treating each segment in isolation.
A third model, Flora-Rewarder, supplies scores that can serve as training feedback. Depending on the instruction it receives, Nvidia says it can rate a complete attempt from 1 to 5, classify each half-second step as normal, suboptimal or failure, or assign scores using scoring rules it defines.
Nvidia estimates that Flora can label video at up to 125 times lower cost than OpenAI’s GPT-6 Astra. Flora’s cost figure is a projection, not a measured production cost: Nvidia calculated it using an assumed rental price of $10 per graphics processing unit (GPU) per hour.
The W27 keeps the Galaxy Z Fold 8 Ultra’s core hardware, with dark-red or dark-blue finishes, golden accents and 16GB of working memory in both storage versions.
2 min
| Model | Annotation quality score | Estimated cost per video hour |
|---|
| Flora | 57.1% | $0.72 |
| GPT-6 Astra | 53.7% | $88.49 |
Nvidia’s reported results vary by test. On the public WGO-Bench benchmark, Flora leads at finding action boundaries but trails GPT-6 Astra overall and Google’s Gemini 3.8 Flash on description accuracy. The researchers attribute Flora’s gains mainly to locating actions accurately in time.
On Nvidia’s DenseCaptionBench, the company reports that Flora leads across the measures for human recordings but ranks second behind GPT-6 Astra for robot videos. The researchers acknowledge an evaluation limitation: GPT-6 Astra judges descriptions while also being one of the models under test.
Nvidia also reports stronger detection of failed steps within a task. That comes with false alarms: Flora-Rewarder incorrectly flags about 11% of normal steps as errors. GPT-6 Astra still leads on scoring against defined criteria and ranking complete attempts.
| Model | Share of failed steps detected |
|---|---|
| Flora-Rewarder | 53.3% |
| GPT-6 Astra | 0% |
| Other tested models | No more than 11.7% |
The researchers' examples also show Flora splitting longer actions into too many segments and sometimes describing them incorrectly. In one video, it labels a robot wiping a plate with a cloth as folding a napkin.
Nvidia says it already uses Flora during training of its GR00T robotics models, annotating up to 200,000 hours of video per day across 512 GPUs. With the models released on Hugging Face, researchers can generate labels and reward scores for their own robotics datasets and test approaches such as transferring human demonstrations to robots.
The satellite operator expects access to more than twice OneWeb’s current bandwidth in return for spending in 2027–2034, with initial IRIS² services expected from 2030.
3 min
The black, gold and green edition follows September’s sold-out preorders. Nintendo’s replenished launch-day stock also quickly sold out.
1 min