What happened
Nvidia released Flora, an open-source data engine for annotating robot video, available on Hugging Face. The system uses three fine-tuned Qwen3.5 models to segment actions, describe them, and evaluate task success. Nvidia reports that Flora costs 72 cents per hour of video to run, compared to approximately $88.50 for OpenAI's GPT-6 Astra, while scoring slightly higher in a 50-video test.
Nvidia has released Flora, an open-source data engine designed to annotate robot video footage. The models are available on Hugging Face, and Nvidia states that Flora is already integrated into the training process for its GR00T models, where it processes up to 200,000 hours of video daily using 512 GB300 GPUs.
The system consists of three fine-tuned Qwen3.5 models. A 9B model segments video into actions, a 27B model generates descriptions for each segment using context, and a third 27B model, Flora-Rewarder, evaluates task success and step-by-step progress. This architecture allows Flora to handle three types of annotations: action timing, step-by-step progress scoring, and task success determination.
Nvidia claims that running Flora costs 72 cents per hour of video, based on an assumed GPU rental rate of $10 per hour. In comparison, using OpenAI's GPT-6 Astra for the same task costs approximately $88.50 per hour. In a 50-video test, Flora scored slightly better than Astra, which was allowed to use its full reasoning capacity. Claude Opus 5 performed worse on the same test at a cost of about $42 per hour.
Flora demonstrates particular strength in detecting failures, with its Flora-Rewarder component catching 53.3 percent of failed steps in episodes, compared to a maximum of 11.7 percent for rival models and zero for Astra. However, this comes with a trade-off of false alarms, as Flora marks about 11 percent of normal steps as errors. On the RoboRewardBench, Flora's error rate is 27.8 percent lower than Astra's, largely due to better handling of partially finished tasks.
Why it matters
Flora addresses the high cost of labeling large-scale robotics datasets, a critical bottleneck in training physical AI models. By offering a free, open-source alternative that handles multiple types simultaneously, it lowers the barrier for researchers and developers to build and refine robot learning pipelines. The significant cost reduction, even if based on projected rates, makes large-scale video annotation economically viable for teams that previously relied on expensive proprietary APIs or manual labeling services.
The primary value of Flora lies in its cost efficiency and open-source nature. Robotics teams often work with massive volumes of video data, making the high per-hour costs of proprietary large language models prohibitive. By offering a tool that is both cheaper and open-source, Nvidia enables researchers to scale their data efforts without incurring significant API fees.
Flora's ability to handle multiple types in a single pipeline simplifies the data preparation process. Previously, teams might have needed different tools for action segmentation, description, and reward modeling. This integration can streamline the workflow for training physical AI models, which rely heavily on labeled video data to learn from human demonstrations and robot interactions.
The release also fits into Nvidia's broader strategy of supporting the entire robotics value chain. By providing tools for data processing and , Nvidia encourages developers to use its hardware for training and , potentially driving demand for its GPUs. This move competes with proprietary services like Scale AI's Physical AI Data Engine and aligns with efforts by other companies, such as Google, to automate the creation of robotics training data.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
Which component of an AI application is the machine-learning model itself?
What to watch next
Monitor independent benchmarks to verify the cost and performance claims, as the initial tests were conducted by Nvidia and used its own hardware. Watch for adoption by robotics labs and whether the open-source release accelerates the development of general-purpose robot models. Also observe how competitors like Google and Scale AI respond to this open alternative in the physical AI data pipeline.
Independent verification of the cost and performance metrics is crucial. The current benchmarks were conducted by Nvidia, which sells the hardware used to run Flora, and the cost estimates are based on projected rental rates rather than actual production bills. Third-party tests will provide a clearer picture of the real-world efficiency and accuracy of the system.
Adoption by the robotics research community will determine the long-term impact of Flora. If researchers find that the open-source tool is reliable and easy to integrate, it could become a standard component in the development of new robot models. This could accelerate progress in physical AI by making high-quality labeled data more accessible.
Competitive responses from other AI companies and data services will be important to watch. Google has recently announced similar capabilities in its Gemini 4 Argon model, and Scale AI offers proprietary solutions. The entry of an open-source alternative from a major hardware provider could shift the market dynamics, potentially leading to further innovation and price competition in the robotics data tooling space.