Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra

The cost of labeling robot training data drops by over 120x, making large-scale robotics development more accessible.

Nvidia has open-sourced Flora, an AI data engine designed to significantly reduce the cost of labeling robot training videos. Flora can reportedly label an hour of robot video for approximately 72 cents, a fraction of the estimated $88.50 cost using OpenAI's GPT-6 Astra. The system is built on three fine-tuned Qwen3.5 models and aims to provide three crucial types of labels: action descriptions for data selection, step-by-step progress ratings for trial-and-error learning, and task completion judgments for deployment monitoring. While most existing tools offer only one label type, Flora addresses all three. Nvidia claims Flora is already annotating up to 200,000 hours of video daily on its GR00B models and has released the models on Hugging Face. Although cost projections are based on assumed GPU rental prices, Nvidia states Flora achieved higher annotation quality than GPT-6 Astra and significantly better mid-episode error detection in tests, though it sometimes misidentifies or splits actions.

Nvidia's open-sourcing of Flora signals a strategic move to bolster its hardware ecosystem by lowering a critical barrier to entry in robotics AI development. By making advanced video annotation dramatically cheaper, Nvidia aims to encourage wider adoption of its GPUs for training sophisticated robotic models, thereby driving demand for its chips. This approach aligns with their broader strategy of dominating the robotics value chain, from hardware to data processing and AI models, potentially creating stickiness for developers within their ecosystem.

The release of Flora challenges commercial data labeling services and suggests a trend towards more accessible, powerful tools for AI training. Its ability to provide multiple label types, particularly for identifying mid-episode errors, could accelerate the development of more robust and reliable robots. Competitors, including Google with its Gemini 4 Argon, are also targeting similar aspects of robot training data, indicating a growing market focus on efficient and automated data processing for robotics advancement.