NVIDIA releases Flora, the three-model data engine it uses to label robot training video

NVIDIA’s GEAR robotics group has published Flora, the system it says labels training data for its GR00T robot models, and put the weights on Hugging Face. NVIDIA researcher Jim Fan announced it on X on 20 October; the paper carries the same date.

What is in it

Flora is three models that turn raw robot or human manipulation video into training signals:

• Flora-TempSeg (9B, built on Qwen3.5-9B) finds where each operation starts and ends in a video.
• Flora-Captioner (27B, built on Qwen3.5-27B) writes one action instruction for each segment, using the footage before and after it for context.
• Flora-Rewarder (27B, built on Qwen3.5-27B) scores whether an episode succeeded and gives dense step-by-step action-quality scores for reinforcement learning.

The model cards say Flora was the data engine used to train GR00T-2. NVIDIA says the production setup labels 200,000 hours of video a day across 512 GB300 GPUs.

The license

The weights are not open source in the usual sense. Each model card says use is governed by the NVIDIA OneWay Noncommercial License, that Flora is “for research and development only,” and that it is “not intended or supported for commercial or production use.” The cards also say Flora should not directly control robots or act as the only safety monitor for a physical system.

Where the 125x comes from

NVIDIA’s headline claim is “Astra-level” annotation quality at about 125 times lower cost than OpenAI’s GPT-6 Astra. The paper spells out the arithmetic. On a fixed 50-episode subset of NVIDIA’s DenseCaptionBench (human, gripper and dexterous-hand footage), Flora scored 57.1% quality against 53.7% for GPT-6 Astra at maximum reasoning effort. Flora’s cost of $0.72 per hour of source video assumes eight GB300 GPUs at $10 per GPU-hour; Astra’s $88.49 per hour comes from API usage priced at public rates from 4 October. The exact ratio is 122.5, rounded to “approximately 125x.” The paper notes that GPU pricing is assumed and that storage, billing granularity and deployment overheads are not included, with a scenario range of $0.51 to $1.26 per hour.

On public benchmarks, NVIDIA reports leading scores on the WGO-Bench captioning test and the RoboReward and Robometer reward tests. Astra is not beaten everywhere: on NVIDIA’s in-house RubricBench, where lower error is better, Flora-Rewarder scored 6.5 against Astra’s 3.6, while beating nine of the ten general-purpose models compared.

Who should look

If you train manipulation policies and pay to label video, Flora is a ready-made pipeline you can run on your own GPUs for research. Before you compare it to an API, swap in your own GPU-hour price. By our arithmetic, at $20 an hour instead of $10 the gap roughly halves, to about 60x, which is still large.