Nvidia opened the Flora source code — a system for labeling videos of robots' actions. According to The Decoder, one hour of such video Flora processes for 72 cents, versus 88 dollars when using GPT-6 Astra, with its quality rating slightly higher.
The system is built on three fine-tuned Qwen3.5 models. In The Decoder's recap it also states that Flora better detects errors in robot actions than competitors, but sometimes lags in the accuracy of descriptions.
For robotics developers this could mean cheaper preparation of training data and an additional open tool. Nvidia ties Flora to the intention to control the entire robotics systems development chain, but the available material does not reveal license terms, the methodology of comparison, or practical deployment results.
editorial commentary
Why it matters
If the claimed figures are confirmed, Flora could reduce the barrier to data preparation in robotics and amplify Nvidia's influence beyond hardware. The next observable signal will be license terms, independent tests, and deployment data. Substantial uncertainty stems from the fact that only a metadata synopsis of one source is currently available.