Nvidia has launched Flora, a model suite designed to automate video labeling for robot training, integrating it into the company’s existing GR00T pipeline. The Flora models became available on Hugging Face on October 25, 2026.
Labeling is one of the slowest steps in robot training. Human reviewers must watch hours of footage, mark actions, and flag mistakes, work that is expensive and hard to scale. Flora aims to do much of that job automatically, annotating up to 200,000 hours of video per day at a significantly lower cost than using a large general-purpose model.
The suite uses a 9-billion-parameter model to segment actions and a 27-billion-parameter model for temporal context, running on 512 GB300 GPUs. In reported tests, it identified 34 of 37 test actions and caught failed steps 53.3 percent of the time, while flagging normal steps incorrectly in about 11 percent of cases.
The system is built from three fine-tuned Qwen3.5 models and is reported to deliver a cost reduction of about 125 times compared with GPT-6 Astra for this labeling work. That comparison comes from the company’s own framing, and the error rates show why human review still matters for safety-critical robots.
For robotics teams, cheaper labeling could mean faster training cycles and more ambitious data collection. But it also ties the workflow to Nvidia’s hardware ecosystem, a trade-off developers will weigh against the savings. Platforms such as Hugging Face make the models easy to find, and local runtimes such as Ollama show how open models increasingly reach developers outside the cloud.
Source: Wisevoter
