Nvidia's Flora labels robot video at 72 cents an hour

Nvidia frames Flora as an open alternative to commercial labeling services such as Scale AI's Physical AI Data Engine. The gap it targets is structural: most existing tools supply only one of the three label types robot training needs, and general models like GPT-6 Astra can produce several but at a cost that does not scale to the video volumes involved. Flora splits the job across specialized models, with a 9B model segmenting video into actions and two 27B models handling description and reward scoring.

The quality picture is mixed. Flora's clearest advantage is temporal localization: on RoboRewardBench its error rate is 27.8 percent lower than Astra's, and its reward model catches 53.3 percent of failed steps within an episode, while no competitor gets above 11.7 percent and Astra catches none. The tradeoff is roughly 11 percent of normal steps falsely flagged as errors. In one example, Flora identified 34 of 37 reference actions while Astra caught four, but Flora sometimes over-splits long actions and misdescribes them.

Flora fits Nvidia's broader push to cover the robotics value chain, from hardware and training data to simulation and models, and to keep developers in its hardware ecosystem. It also adds label and reward signals to the automated data pipeline Nvidia showed at GTC. Google's Gemini 4 Argon targets a similar part of robotics training, suggesting the competition over training data tooling is widening.