Nvidia open-sources Flora to cut robot video labeling costs by 125x

The cost of training robots

Training large AI models for robotics requires massive amounts of video data. To make these models effective, researchers need specific text annotations for every action shown in a video. Without these labels, robots struggle to learn from trial and error or understand the goals of a task.

Currently, many tools only provide one type of label at a time. This forces developers to use multiple systems or expensive commercial models to get a complete picture. As reported by The Decoder, these costs can become prohibitive when trying to process the high volumes of video needed for advanced robotics.

Nvidia's new engine, Flora, aims to solve this by providing three types of labels at once. These include descriptions of actions, ratings of how well a robot progresses through steps, and judgments on whether a task was completed successfully. By bundling these features, Nvidia promises to make labeling one hour of robot video cost just 72 cents.

How Flora processes video

Flora is built using three different fine-tuned Qwen3.5 models that work together to analyze footage. A 9B model is responsible for splitting videos into distinct action segments. A 27B model then describes those actions while considering what happened before and after each specific movement.

The third component is a 27B model called Flora-Rewarder. This model handles three different roles depending on the instructions it receives. It can rate an episode's success on a scale of 1 to 5, classify half-second steps as normal, suboptimal, or failures, and assign scores based on specific rubrics.

Nvidia's research lab, GEAR, has already integrated Flora into the training of its GR00T models. The system is currently capable of processing 200,000 hours of footage every day using a cluster of 512 GB300 GPUs. This scale allows for the rapid creation of high-quality datasets for complex physical tasks.

Performance vs. commercial models

Nvidia compared Flora against several high-end models, including GPT-6 Astra and Claude Opus 5. In a test of 50 videos, Flora scored slightly better than GPT-6 Astra despite the latter having access to its full reasoning capacity. While Astra is often more precise in its descriptions, Flora excels at detecting where actions begin and end.

The cost difference is the most striking finding in the study. Nvidia estimates that Flora costs about 72 cents per hour, while GPT-6 Astra costs roughly $88.50 per hour. In the same evaluation, Claude Opus 5 performed significantly worse than the others and still required about $42 per hour. These figures are projections based on a hypothetical GPU rental rate of $10 each per hour.

Flora also shows a specific advantage in catching errors during partially completed tasks. For example, if a robot tips a bottle over instead of knocking it off a table, Flora correctly assigns a partial score. In the same scenario, Astra gave a much lower score because it failed to recognize the partial success. Flora caught 53.3 percent of failed steps, while no other competitor reached 11.7 percent.

Open source and market impact

Nvidia is positioning Flora as an open-source alternative to expensive commercial services like Scale AI's Physical AI Data Engine. By making the tool public, Nvidia allows researchers to scale their own robotics datasets and test new methods for transferring human demonstrations to robots. The models are currently available for the public on Hugging Face.

This move is part of a broader strategy by Nvidia to invest billions in open AI models. By providing the tools to create training data, Nvidia aims to keep developers within its hardware ecosystem. This creates a pipeline where developers use Nvidia chips to run the models that Flora helps train.

The company has been working to solve the shortage of robotics training data through simulation and automated processing. Flora adds a layer of detailed labels and reward signals to this existing pipeline. This helps Nvidia cover the entire robotics value chain, from the physical hardware to the data and the final AI models.