Nvidia's Flora Open-Source Engine Slashes Robot Video Annotation Costs by 125x Compared to GPT-6 Astra

Nvidia's GEAR research lab has open-sourced Flora, a robot video annotation engine that labels training data up to 125 times cheaper than frontier models like GPT-6 Astra, making advanced video annotation accessible at an estimated cost of $0.72 per hour of video. For broader context, explore our Top 100 AI Tools.

Flora is already in production, supporting the training of Nvidia's GR00T robot models. It can annotate up to 200,000 hours of video daily across 512 GB300 GPUs, demonstrating its capacity for large-scale data processing.

Flora's Cost-Efficiency and Performance

Nvidia's Flora engine offers a significant cost advantage in robot video annotation. The company estimates that labeling one hour of video with Flora costs approximately $0.72. This figure is based on an assumed $10-per-GPU-hour rental price for eight GB300 GPUs. In contrast, labeling the same hour of video with GPT-6 Astra is projected to cost around $88.50, making Flora up to 125 times more cost-effective.

Beyond cost, Flora also demonstrated competitive performance in Nvidia's 50-video quality test, scoring slightly higher than GPT-6 Astra. Specifically, Flora achieved 57.1 percent annotation quality at its $0.72 per video hour cost. This compares to GPT-6 Astra's 53.7 percent at $88.49 per hour and a lower score for Claude Opus 5, estimated at about $42 per hour.

Under the Hood: Flora's Architecture

Flora is built upon three fine-tuned Qwen3.5 models, each specialized for different aspects of video annotation:

• 9B Segmentation Model: This component is responsible for segmenting video content.
• 27B Captioning Model: This model generates action descriptions for the video segments.
• 27B "Flora-Rewarder" Model: This specialized model provides step-by-step progress ratings and success judgments for robotic actions.

The Flora-Rewarder model is particularly notable for its ability to identify failed steps mid-episode, catching 53.3 percent of such instances. This capability significantly surpasses competitors, which do not exceed 11.7 percent in similar tasks.

Availability and Use Cases

The Flora models are openly available on Hugging Face, providing researchers, academics, and the open-source community with tools for annotating robot datasets and exploring emerging applications like physical auto-research and human-to-robot transfer. Developers can also use Flora for prototyping and evaluating data curation, filtering, and monitoring pipelines in robotics foundation model research.

The open-sourcing of Flora provides a practical solution for reducing the financial and computational barriers associated with training advanced robotics models. Its integration into Nvidia's GR00T training pipeline underscores its utility and scalability in real-world applications.

Conclusion

Nvidia's release of Flora as an open-source robot video annotation engine marks a significant step in making high-quality robotics training data more accessible and affordable. By drastically cutting annotation costs while maintaining competitive quality, Flora enables broader research and development in robotics. Its availability on Hugging Face and its proven performance in production environments suggest it will play a crucial role in advancing the next generation of AI-powered robots.

Sources

• Flora - a nvidia Collection
• GitHub - NVIDIA/Flora: The open-source model suite for robotics data engine by NVIDIA GEAR · GitHub
• nvidia/Flora-Captioner · Hugging Face
• nvidia/Flora-TempSeg · Hugging Face
• README.md · nvidia/Flora-Captioner at main