The Pulse
NVIDIA Releases Flora Models for Robot-Video Labeling
NVIDIA has released Flora, a three-model suite that segments, describes and scores human and robot manipulation video. The company says the system can label 200,000 hours of data per day, but its model cards restrict use to noncommercial re
AI.info Team · Written with AI · fact-checked ·
“Flora is battle tested at scale, labeling 200,000 hours of data per day across 512 GB300s.”
Sicong Leng et al., Flora
NVIDIA’s Flora models turn human and robot manipulation footage into labels that can feed robotics training and evaluation. The company’s model cards list the weights as released on October 15, 2026; its research paper is dated October 20. NVIDIA says the suite is already used to train its GR00T models and is intended to help researchers build custom robotics datasets.
Flora does not control a robot. Instead, it annotates video with action boundaries, language descriptions and reward scores—signals used to organize training data, assess task progress and monitor robot performance. NVIDIA’s project page presents the release as an open data engine, while its model card limits use to noncommercial research and development.
Three models turn footage into labels
Flora divides annotation among three models. TempSeg identifies when actions begin and end in a recording; Captioner writes an instruction for each selected interval; Rewarder scores outcomes or actions within an episode. Researchers can also supply a rubric to define the criteria for a reward score.
The pipeline addresses different stages of robot learning. Timestamped captions can help filter and balance pretraining data, while process rewards measure progress during reinforcement-learning post-training. Outcome rewards can support evaluation of completed tasks or monitoring of deployed systems, according to NVIDIA.
The released weights are available through NVIDIA’s Flora collection on Hugging Face. The Captioner model card says the model was developed for GR00T-2 training and for researchers building custom robot datasets; it also states that the output consists of annotations, not robot commands.
200,000 hours is a production-scale figure
NVIDIA reports throughput of up to 200,000 hours of video per day across 512 GB300 GPUs. That is the company’s production-scale figure, not a claim about what one workstation can process. The paper also compares annotation cost on a separate test: a 50-episode subset of its DenseCaptionBench evaluation.
For that comparison, NVIDIA estimates Flora’s cost at 72 cents per source-video hour using eight GB300 GPUs priced at $10 per GPU-hour. It puts GPT-6 Astra’s API cost at $88.49 per video hour at maximum reasoning effort. On the same comparison, Flora scores 57.1% on the paper’s combined annotation-quality measure, against Astra’s 53.7%; NVIDIA calculates an approximately 125-fold cost difference from those estimates.
The cost result depends on the paper’s hardware and API-cost assumptions, rather than a direct measurement of every lab’s operating costs. NVIDIA’s paper describes the result as an estimate and points readers to its appendix for the accounting details.
Benchmark results vary by task
Flora does not lead every comparison. On WGO-Bench, which includes 100 human and robot manipulation episodes, the paper reports 50.3% for Flora on segmentation F1 and 64.7% on gold-window caption accuracy. GPT-6 Astra scores higher on caption accuracy, at 77.4%, and on joint semantic F1, at 39.0% versus Flora’s 34.5%.
Results also differ between human and robot footage in NVIDIA’s DenseCaptionBench tests. Flora leads the listed models across all three reported Human100 measures, including a 37.4% joint point F1. On Robot200, Astra scores 57.9% on joint point F1, ahead of Flora’s 51.0%.
Reward tests show a similar mix of strengths. Flora records a mean absolute error of 40.0 on RoboRewardBench, compared with 55.4 for Astra, and ranks robot trajectories with a Kendall tau score of 72.0 on the out-of-distribution Robometer test, below Astra’s 82.3.
Open weights come with a noncommercial limit
The Captioner card names NVIDIA’s OneWay Noncommercial License and says Flora is intended for noncommercial robotics research and development, not commercial or production use. That restriction narrows who can adopt the weights directly: a research lab can test the models under the stated terms, but a company seeking production annotation cannot treat the release as unrestricted commercial software.
The paper also identifies a concrete weakness in failure scoring. On DenseRewardBench, Flora detects 53.3% of labeled failure steps, but it also marks about 11% of normal steps as failures. The scores may help researchers inspect robot footage; they do not establish reliable safety monitoring on their own.