Nvidia Open-Sources Flora to Label Robotics Videos for 72 Cents per Hour

Nvidia has announced making the Flora data engine available as open source, stating that it reduces the cost of labeling one hour of robotics video to around 72 cents, compared to about $88.50 when using OpenAI's GPT-6 Astra. Robotics models rely on videos requiring text annotations to learn motion and evaluate performance.

Flora consists of three modified Qwen3.5 models. A 9-billion parameter model segments the video into motion clips, a 27-billion parameter model describes each motion considering what happened before and after it, and a second model of the same size called Flora-Rewarder evaluates the episode's success and gives scores to each step. The system covers three types of annotations needed for robotics training.

Nvidia said the system is already working during the training of its GR00T models, labeling up to 200,000 video hours per day on 512 GB300 units, and that the models are available on the Hugging Face platform. In a test on 50 videos, Flora achieved a labeling quality of 57.1% compared to 53.7% for GPT-6 Astra, but cost figures remain estimates rather than production measurements, as they were calculated based on a hypothetical rental price of $10 per GPU hour.

Results showed Flora's superiority in detecting motion boundaries and identifying errors within an episode, capturing 53.3% of failed steps compared to GPT-6 Astra not capturing any steps, though it mistakenly labels about 11% of normal steps as errors. On the general WGO-Bench benchmark, Astra leads in overall score and advances in pure description accuracy on robotics videos. Nvidia offers Flora as an open alternative to commercial services like Scale AI's physical AI data engine.

What do these terms mean?

Data labeling: Adding descriptions and evaluations to video clips or text to train an AI model to understand them, which is considered one of the largest costs in training robotics models.

Parameters: Adjustable numerical values within a model, their count being related to the model's size and capability, reaching 9 billion and 27 billion here.

Graphics Processing Units (GPUs): Chips designed to execute parallel computational operations efficiently, widely used to train and run AI models.