- Nvidia open-sourced Flora, a video-labeling system built from three fine-tuned Qwen3.5 models, to generate the action descriptions and reward scores robot-training videos need.
- By Nvidia’s own estimate, Flora labels one hour of robot video for about 72 cents, versus roughly $88.50 for GPT-6 Astra and about $42 for Claude Opus 5 in the same test.
- The system is already running in production, annotating up to 200,000 hours of video per day across 512 GB300 GPUs to train Nvidia’s GR00T robot models.
- On Nvidia’s RoboRewardBench, Flora’s error rate on mid-episode failures came in 27.8% lower than GPT-6 Astra‘s, though GPT-6 Astra still led on rubric-based scoring and the public WGO-Bench.
What Happened
Nvidia’s GEAR research lab open-sourced Flora on October 25, 2026, an open data engine that labels robot-training videos with action descriptions, step-by-step progress ratings, and success-or-failure judgments. The models are available on Hugging Face.
Why It Matters
Training large robot models requires annotating enormous volumes of video, and general-purpose models like GPT-6 Astra can do the job but are too expensive to run at scale. Flora is designed specifically to cut that cost while keeping the labeling pipeline open rather than tied to a commercial service like Scale AI’s Physical AI Data Engine.
Technical Details
Flora pairs a 9B segmentation model that splits video into action segments with a 27B captioning model and a second 27B model, Flora-Rewarder, that scores episode success on a 1-to-5 scale and classifies each half-second step as normal, suboptimal, or failure. In a 50-video test, Nvidia says Flora scored slightly better overall than GPT-6 Astra and caught 53.3% of failed steps within an episode, versus no more than 11.7% for any competing model it tested. The tradeoff: Flora falsely flags about 11% of normal steps as errors, and it trails GPT-6 Astra on the public WGO-Bench and on raw description precision against Gemini 3.8 Flash.
Who’s Affected
Robotics researchers who need annotated training data at scale get a free alternative to expensive proprietary labeling. Nvidia also benefits directly: Flora feeds its GR00T training pipeline, building on the simulated training environments Nvidia has already built for robotics, and reinforces demand for its own GPUs, which the cost comparison was built on in the first place. The release follows Nvidia’s recent Nemotron 3.5 Bolt launch as part of a broader push into open, locally-deployable AI tooling.
What’s Next
Nvidia says Flora’s gains are concentrated in detecting action boundaries and timing, areas it plans to keep refining, while Google’s newly announced Gemini 4 Argon is targeting a similar slice of robot-training video analysis — meaning the labeling-cost comparisons are unlikely to stay settled for long.
