Nvidia Flora AI Automates Robot Video Labeling Tasks

The new model slashes data annotation costs for robotics developers by optimizing video parsing on Nvidia's GB300 hardware.

Updated on Oct. 26, 2026 in Robotics

Isometric editorial illustration of a modular robotic arm on a clean surface, representing automated data processing for robotics research. — Nvidia has launched its Flora AI model, which automates video data labeling for robotics training on the company's GB300 GPU infrastructure. AI Illustration. Upload story photo >

Live Poll

Is open-source AI development more beneficial for the public than proprietary corporate AI models?

Nvidia has released its Flora AI model, designed to automate video labeling for robot training at a fraction of existing costs. The system is now integrated into Nvidia's GR00T robot learning pipeline and is publicly available on Hugging Face.

Why it matters

By dramatically lowering the barrier to high-quality data annotation, Nvidia aims to accelerate robot model training while cementing demand for its high-performance GB300 GPU infrastructure.

Flora labels 200,000 hours of video daily using 512 GB300 GPUs, achieving a 72-cent hourly cost that is approximately 125x cheaper than GPT-6 Astra. In internal testing, the system identified 34 of 37 action steps, catching 53.3 percent of failures while maintaining an 11 percent false-positive rate.

The players

Nvidia

A dominant designer of accelerated computing chips and software stacks, currently positioning its GB300 hardware and GR00T framework as the standard for physical AI.

Hugging Face

An open-source hub for machine learning models that hosts the Flora suite for developer access.

The details

Flora uses three fine-tuned Qwen3.5 models to parse spatial and temporal data. A 9B-parameter model breaks video into discrete action segments, while a 27B-parameter model interprets these segments by factoring in the surrounding temporal context. This dual-model architecture allows the system to identify failed steps in an episode, automating a task that traditionally requires intensive human oversight.

Earlier this year, Nvidia demonstrated its foundational data pipeline at GTC.
October 25, 2026: Flora models were released on Hugging Face.

The Tech Race

The release of Flora marks the commercial realization of the data annotation infrastructure Nvidia first teased at its GTC event earlier this year. It positions Nvidia directly against general-purpose LLM providers by proving that purpose-built, smaller models can outperform massive, expensive alternatives in specialized robotics workflows.

Developers and researchers can access the Flora models immediately via Hugging Face to integrate automated labeling into their own robotics training stacks. While the models are open for use, deploying them at scale requires significant investment in Nvidia's GB300 GPU infrastructure.

The takeaway

Nvidia's move to commoditize high-speed video labeling effectively lowers the cost of entry for building capable physical robots. Developers should monitor upcoming robot benchmark results to see if Flora's 11 percent false-flag rate hinders performance in complex multi-step tasks.