Nvidia's Flora labels robot video for an estimated 72 cents an hour

Nvidia open-sourced Flora, a suite of video-labeling models, and put a number on what the industry normally pays to turn raw footage into training data: about 72 cents an hour of video, against roughly $88.50 with a frontier model.

Flora comes from Nvidia's GEAR research lab and is built to do the three labeling jobs robot training needs. A 9-billion-parameter model splits a video into action segments. A 27-billion-parameter model writes a plain-language instruction for each segment and weighs what happened before and after it. A second 27-billion-parameter model, Flora-Rewarder, scores whether a step worked. All three are fine-tuned from the Qwen3.5 family and are published on Hugging Face with the paper.

The cost figures are Nvidia's projections, not measured bills. The company modeled an eight-GB300 deployment at $10 per GPU-hour and estimated $0.72 per hour of source video for Flora, against $88.49 for GPT-6 Astra and about $42 for Claude Opus 5. That is a gap of more than a hundredfold, and the company states the assumption rather than hiding it.

Quality is closer than the price suggests. On a 50-video test, Flora scored 57.1% on annotation quality against 53.7% for GPT-6 Astra, and Nvidia ran most of that testing itself. Where Flora stood out was finding actions: in a clip built from short hand movements, it identified 34 of 37 individual actions while Astra caught 4. Its 53.3% catch rate for failed steps beat every competitor in the test, none of which topped 11.7%. The trade-off is false alarms: Flora wrongly flags about 11% of normal steps.

Nvidia researcher Jim Fan wrote on X that Flora already runs inside the company's own data engine, labeling 200,000 hours of video a day on 500 GPUs. Nvidia positions the suite as an open alternative to commercial pipelines such as Scale AI's Physical AI Data Engine.

The upside case is straightforward: every robotics team that trains on video is paying for annotations on every hour it collects, and the price is the constraint on how much data a model ever sees. Cutting that cost by two orders of magnitude buys a lot more labeled hours. The downside is that a cheap labeler with an 11% false-positive rate can teach a robot the wrong lesson at scale, and quality claims from the vendor's own test are the softest kind.

What to watch is whether anyone outside Nvidia reports Flora's labels holding up on their own footage. The models are downloadable now.