Nvidia Makes Robot Video Labeling 125x Cheaper

Nvidia is the company known for its chips and AI work. It has released the code for Flora, a system that automatically labels robot videos for training AI models. Flora already processes up to 200,000 hours of video a day. Inside it are three smaller models built on Qwen3.5, a family of AI models. One model splits each clip into separate actions. Another describes those actions in words. The third judges whether the robot actually completed the action or not. By Nvidia's own numbers, this labeling costs about 125 times less than GPT-6 Astra, a large AI model used for the same task.

Flora is a narrow tool for labeling robot training video. Most AI agent developers will not need it directly. What matters here is not Flora itself but the result. A small model built for one specific job beat GPT-6 Astra, and not just on price. It also scored higher on accuracy: 54.6 versus 44.2 on the action-splitting benchmark. This is one more case where the biggest model is not always the right tool. Sometimes a narrow, cheap model just does the job better.