Nvidia opens Flora with video notes for robots at 72 cents an hour

Nvidia released Flora, an open set of 3 AI models that make text notes for video to train robots. In a test with 50 videos, Nvidia tells that Flora gave a result that is more correct than GPT-6 Astra, by a small difference. Nvidia estimates 72 cents for one hour of video with Flora and $88.50 with Astra, with a GPU rental price of $10 each hour.

How it works

Robot models train on video with text notes of 3 types:

• the steps of the robot, with times
• how good each step is
• if the task is complete

Most tools give only 1 type. Flora gives all 3, with models made from Qwen3.5. Nvidia tells that it uses Flora at this time. Flora makes text notes for a maximum of 200,000 hours of video each day for GR00T robot models.

The numbers

Flora finds 53.3 percent of the steps with an error. No other model finds more than 11.7 percent, and Astra finds none. On the RoboRewardBench test, its error rate is 27.8 percent lower than the Astra rate.

What is not known

Flora marks about 11 percent of usual steps as errors. In some tests Astra gives more accurate text for robot video. In one test, Astra checked the text and was also one of the models in the test.

Insight: Nvidia gives away software to keep demand for GPUs

In October 2026 Nvidia also released open Nemotron models for agent document search and open models at GTC Berlin. Flora adds robot video notes. The source tells that Nvidia makes open models to keep developers on its hardware. This reads as a toll gate: free tools at each step, and all work ends at Nvidia hardware. If you use free Nvidia tools, check which GPUs they use.