Nvidia open-sources Flora, a cheaper robot video labeling engine

Large robot models need three kinds of labels on training video: descriptions of actions, step-by-step ratings of progress, and judgments on whether a task succeeded. According to Nvidia's GEAR research lab, most existing tools supply only one of these, while large language models such as GPT-6 Astra can produce several but cost too much at scale. Flora is built from three fine-tuned Qwen3.5 models: a 9B model that splits video into action segments, a 27B model that describes them, and a 27B reward model called Flora-Rewarder.

In a test with 50 videos, Nvidia says Flora scored slightly better than GPT-6 Astra, which was allowed its full reasoning capacity, reaching 57.1 percent annotation quality against 53.7 percent. The cost figures assume a rental price of $10 per GPU per hour and are not production measurements. Nvidia says Flora already annotates up to 200,000 hours of video per day while training its GR00T models, and the models are available on Hugging Face.

Results are mixed. Astra wins on the WGO-Bench overall score, robot videos and rubric scoring, and the researchers acknowledge that Astra also grades descriptions while being one of the models under test. Flora catches 53.3 percent of failed steps within an episode, where Astra catches none, but it falsely flags about 11 percent of normal steps as errors.

Developers building robot agents can get action labels, step ratings and failure detection from an open model instead of paying for large language model annotation. Flora misses some actions and sometimes gets descriptions wrong, so its output should be checked before it feeds training or deployment monitoring. The cost comparison rests on Nvidia's own estimates.