Nvidia’s Flora Offers Cost-Effective Robot Video Labeling Compared to GPT-6 Astra
Nvidia’s new data engine, Flora, provides a cost-effective solution for labeling robot video, charging just 72 cents per hour compared to GPT-6 Astra’s $88. However, the comparison is nuanced, as the quality differences are not as stark as the price suggests.
Large AI models for robotics require annotated videos, which Nvidia’s GEAR research lab identifies as needing three types of labels: action descriptions, step-by-step progress ratings, and success judgments for task completion. Most existing tools offer only one type, while large language models like GPT-6 Astra can generate multiple labels but at a prohibitive cost for extensive video datasets.
Flora effectively segments videos into time intervals, describes actions, and assesses the execution and results of each episode. According to Nvidia’s research, Flora encompasses all three labeling types and is currently utilized in training its GR00T models, annotating up to 200,000 hours of video daily across 512 GB300 GPUs. The models are accessible on Hugging Face.
Flora consists of three fine-tuned Qwen3.5 models. A 9B model segments videos into action parts, while a 27B model describes these actions considering prior and subsequent events. Another 27B model, Flora-Rewarder, evaluates episode success on a scale of 1 to 5, classifies actions as normal, suboptimal, or failures, and assigns scores based on custom criteria.
Nvidia estimates that labeling one hour of video with Flora costs about 72 cents, compared to approximately $88.50 with GPT-6 Astra. In tests involving 50 videos, Flora reportedly outperformed GPT-6 Astra slightly, despite the latter being allowed to utilize its full reasoning capabilities.
Flora achieves an annotation quality of 57.1 percent at $0.72 per hour, while GPT-6 Astra reaches 53.7 percent at $88.49. However, these cost figures are projections based on an assumed rental price of $10 per GPU per hour.
In terms of performance, Flora excels at detecting action boundaries but does not match Astra’s overall score and falls behind Gemini 3.8 Flash in descriptive accuracy. The improvements in Flora primarily stem from its temporal localization capabilities.
On Nvidia’s DenseCaptionBench, Flora leads in metrics for human recordings but ranks second behind Astra for robot videos. This discrepancy arises partly because GPT-6 Astra evaluates descriptions while also being tested.
Flora’s error rate is 27.8 percent lower than Astra’s on RoboRewardBench, attributed to its ability to recognize partially completed tasks. For example, if a robot tips over a bottle instead of knocking it off a table, Flora assigns a higher score to reflect human judgment, whereas Astra assigns a much lower score.
Flora-Rewarder detects 53.3 percent of failed steps within an episode, while no competitor exceeds 11.7 percent, and Astra fails to identify any. However, Flora does misidentify about 11 percent of normal steps as errors. Astra still leads in rubric-based scoring and trajectory ranking.
In a specific test, Flora identified 34 out of 37 actions in a video, while Astra only recognized four. Nonetheless, Flora sometimes divides longer actions into excessive small steps and can misdescribe them; for instance, it inaccurately described a robot’s action of wiping a plate as folding a napkin.
Nvidia positions Flora as an open alternative to commercial services like Scale AI’s Physical AI Data Engine, enabling researchers to scale their robotics datasets and explore methods for transferring human demonstrations to robots. Nvidia aims to keep developers within its hardware ecosystem and boost demand for its chips through significant investments in open AI models, with Flora serving this strategy.
The company has also been addressing the shortage of robotics training data through simulation and automated data processing. Earlier this year at GTC, Nvidia showcased a data pipeline that automatically prepares, multiplies, and scores training materials. Flora enhances this pipeline by adding detailed labels and reward signals from video.
This initiative aligns with Nvidia’s broader strategy to dominate the robotics value chain, covering hardware, training data, simulation, and AI models. Google’s recently announced Gemini 4 Argon targets similar aspects of the robotics training process, identifying work steps in training videos and flagging failed actions.
Source: The Decoder