Nvidia 开源机器人视频标注引擎 Flora，称成本比 GPT-6 Astra 低 125 倍

Labeling one hour of robot video for 72 cents instead of $88 is what Nvidia promises with Flora, its open data engine. But the comparison with GPT-6 Astra isn't quite as lopsided as the price tag suggests.

Large AI models for robots learn from videos that need text annotations, and according to Nvidia's GEAR research lab, this requires three types of labels. First, descriptions of what action happens when in the video, used to select training data. Second, ratings of how well a robot progresses step by step, so it can learn through trial and error. And third, judgments on whether a task was completed successfully, used to monitor robots during deployment. Most existing tools only provide one of these types. Large language models like GPT-6 Astra can generate several, but they're too expensive for the massive volumes of video involved.

alt: Overview graphic of Flora's five capabilities, including a robot stacking cups into a pyramid, a timeline with segments A through G, example captions, and examples of outcome, process, and rubric rewards. — Flora breaks videos down into time segments, describes each action, and evaluates the execution and outcome of an episode. | Image: Nvidia

According to the paper, Flora covers all three. Nvidia says the system is already running during training of its GR00T models, annotating up to 200,000 hours of video per day across 512 GB300 GPUs. The models are available on Hugging Face.

Flora consists of three fine-tuned Qwen3.5 models. A 9B model splits videos into action segments, and a 27B model describes them while 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.

alt: Diagram of the Flora pipeline showing an episode video branching into timestamped captions via overlapping windows and into reward labels with dense, outcome, and rubric scores. — A segmentation model and a captioning model produce timestamped descriptions, while a single reward model generates all three types of ratings. | Image: Nvidia

Up to 125x cheaper by Nvidia's own estimates

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. Labeling one hour of video runs about 72 cents with Flora by Nvidia's estimate, compared to roughly $88.50 with GPT-6 Astra. Claude Opus 5 scored much worse in the same test and still cost about $42 per hour.

alt: Scatter plot comparing annotation quality in percent against cost per source video hour in US dollars on a logarithmic scale, with Flora in the upper left ahead of GPT-6 Astra, Claude Opus 5, Gemini, and other models. — Flora hits 57.1 percent annotation quality at $0.72 per video hour according to Nvidia, while GPT-6 Astra reaches 53.7 percent at $88.49. | Image: Nvidia

The Flora cost figures are projections, though, not production measurements. Nvidia calculated them based on an assumed rental price of $10 per GPU per hour.

Flora finds actions better, but Astra often describes them more precisely

On the public WGO-Bench, Flora is best at detecting action boundaries but trails Astra in overall score and falls behind Gemini 3.8 Flash in pure description accuracy. According to the researchers, the gains come mainly from temporal localization.

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

alt: Six bar charts comparing Flora with GPT-6 Astra, GPT-5.6 Sol, Gemini 3.8 Flash, and Kimi K3 on timestamped captions and reward annotations. — Flora leads on human recordings, outcome scoring, and step-by-step ratings, while GPT-6 Astra wins on WGO-Bench, robot videos, and rubric scoring. | Image: Nvidia

Almost no other model catches mid-episode errors

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, while no competitor gets above 11.7 percent and Astra doesn't catch a single one. The tradeoff is that Flora falsely flags about 11 percent of normal steps as errors. Astra still leads on rubric-based scoring and trajectory ranking.

In a video with many short hand movements, Flora identifies 34 of 37 individual actions, while Astra catches only four. On the flip side, Flora sometimes splits longer actions into too many small steps and occasionally gets descriptions wrong. In one video, a robot hand wipes a plate with a cloth, but Flora describes it as folding a napkin.

alt: Comparison of Flora and GPT-6 Astra with timelines of detected actions, video stills of hands wiring and passing tea leaves, and examples of a failed toy airplane assembly and a tipped bottle. — In one example episode, Flora identifies 34 of 37 reference actions, while GPT-6 Astra catches only four. | Image: Nvidia

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. 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 company has also been working for some time 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.

All of this feeds into Nvidia's strategy of covering the entire robotics value chain, from hardware and training data to simulation and the AI models themselves. 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.