Nvidia has released Flora, an open-source SI data engine designed to label robot training videos, claiming it reduces costs by up to 125 times compared to using GPT-6 Astra. The move aims to address the bottleneck of high-quality training data for robotics SI models.
What Happened
Flora is developed by Nvidia's GEAR research lab and consists of three fine-tuned Qwen3.5 models. A 9B model segments videos into actions, while a 27B model generates descriptions based on temporal context. A second 27B model, Flora-Rewarder, evaluates episode success on a 1-to-5 scale, classifies steps as normal, suboptimal, or failure, and assigns scores based on custom rubrics. According to Nvidia, the system is already annotating up to 200,000 hours of video per day across 512 GB300 GPUs during the training of its GR00T models.
The company estimates the cost at approximately 72 cents per hour of video, compared to roughly $88.50 for GPT-6 Astra. These figures are projections based on an assumed GPU rental price of $10 per hour. In internal tests, Flora reportedly achieved 57.1 percent annotation quality versus 53.7 percent for GPT-6 Astra, though the latter was allowed to use its full reasoning capacity.
Why It Matters
For developers building robotics SI agents, the cost of data labeling is a significant barrier. While Flora offers a cheaper alternative, its performance is nuanced. On the public WGO-Bench, Flora excels at detecting action boundaries but trails GPT-6 Astra in overall score and falls behind Gemini 3.8 Flash in description accuracy. A methodological caveat noted by researchers is that GPT-6 Astra was used to evaluate the descriptions while also being a model under test.
Flora shows particular strength in error detection. On RoboRewardBench, its error rate was 27.8 percent lower than Astra's, primarily because it better recognizes partially completed tasks. It caught 53.3 percent of failed steps within an episode, whereas competitors scored below 11.7 percent. However, this sensitivity comes with a tradeoff: Flora falsely flags about 11 percent of normal steps as errors. Nvidia positions Flora as an open alternative to commercial services like Scale AI's Physical SI Data Engine, reinforcing its strategy to keep developers within its hardware ecosystem.
The Bottom Line
Nvidia's Flora provides a low-cost, open-source option for robot video annotation, particularly strong in temporal localization and error detection. However, it does not uniformly outperform commercial SI models like GPT-6 Astra, which remain superior in certain descriptive and rubric-based scoring tasks. The release highlights the ongoing industry effort to scale robotics training data efficiently.