Nvidia’s research arm GEAR has open‑sourced a system called Flora that adds three kinds of labels to robot videos: what action is happening, how well each step is performed, and whether the overall task succeeds. The company says the tool can label an hour of footage for roughly 72 cents, compared with about $88.50 using OpenAI’s GPT‑6 Astra. The cost estimate assumes a $10‑per‑GPU‑hour rental price.
Flora is built from three fine‑tuned Qwen 3.5 models. A 9‑billion‑parameter model slices videos into action segments. A 27‑billion‑parameter model writes captions that consider surrounding actions, and a second 27‑billion‑parameter model – called Flora‑Rewarder – produces the success rating, step‑by‑step classification, and rubric scores. Nvidia reports that the system is already annotating up to 200 000 video hours per day on a farm of 512 GB300 GPUs, and the models are posted on Hugging Face.
How Flora compares with other large language models
In Nvidia’s own benchmark of 50 robot videos, Flora achieved a 57.1 % annotation quality score at $0.72 per hour, while GPT‑6 Astra scored 53.7 % at $88.49 per hour. Claude Opus 5 was cheaper than Astra at about $42 per hour but performed much worse. On the public WGO‑Bench, Flora excelled at finding action boundaries but lagged behind Astra and Gemini 3.8 Flash on overall description accuracy. On Nvidia’s DenseCaptionBench, Flora led for human recordings but was second to Astra for robot videos.
Why the cost and capability matters for robotics developers
Labeling large video datasets is a bottleneck for training robot‑control models. By reducing the price per hour of annotation, Flora could let smaller labs and startups build bigger training sets without paying commercial services like Scale AI’s Physical AI Data Engine. The open‑source release also lets developers keep the data pipeline inside Nvidia’s hardware ecosystem, which may drive more purchases of Nvidia GPUs.
Why it matters
For robotics teams, cheaper labeling means they can train more capable models faster and on larger datasets without huge budgets. The open‑source nature lets researchers customize the pipeline, but the cost claims rely on Nvidia’s internal pricing assumptions, so real‑world expenses may differ.
