- Nvidia's GEAR lab released Flora, an open model set that labels robot video. Nvidia estimates 72 cents per hour against $88.50 for GPT-6 Astra.
- The cost figures are projections based on an assumed $10 per GPU-hour. Flora catches 53.3 percent of failed steps and falsely flags about 11 percent of normal ones.
- Teams that adopt cheap automated labeling should measure its miss rate and false-alarm rate on their own footage before trusting it.
The robot's report card is now written by a model
When a robot fails at work, someone has to notice. That someone is increasingly another AI model. It watches the video, decides whether the task succeeded, and writes the verdict into the training data.
Nvidia's GEAR research lab has released a system called Flora that does this job. The idea worth taking from it is simple. Once grading gets cheap, the quality of the grader matters more than the quality of the price tag.
What Nvidia released
Nvidia says robotics models need three kinds of labels. The first describes which action happens when, so teams can select training data. The second rates progress step by step, so a robot can learn by trial and error. The third judges whether a task was completed, so deployed robots can be monitored.
According to Nvidia, Flora covers all three. The lab describes it as a data engine already running at scale: across 512 GB300 chips, it annotates 200,000 hours of footage each day as part of GR00T model training. The models are available on Hugging Face.
Nvidia estimates the cost at about 72 cents per hour of video. It puts GPT-6 Astra at roughly $88.50 for the same hour. That is the basis for the headline claim of about 125 times cheaper.
How it works
Flora is three fine-tuned Qwen3.5 models, according to The Decoder's report on the paper. A 9-billion-parameter model splits video into action segments. A 27-billion-parameter model describes each one, using what happened before and after.
A second 27-billion-parameter model, Flora-Rewarder, does the grading. It rates an episode's success from 1 to 5. It also sorts each half-second step into normal, suboptimal or failure. And it scores work against rubrics it defines itself.
What the numbers show, and what they do not
The cost figures are projections, not production measurements. The Decoder notes that Nvidia built them on a hypothetical hourly rental rate of $10 for each GPU. Change that rate and the gap changes with it.
The accuracy results are mixed. Nvidia reports a 50-video comparison in which Flora edged ahead of GPT-6 Astra, although Astra was allowed its full reasoning capacity. In the same comparison, Claude Opus 5 did far worse yet was priced at around $42 an hour.
On the public WGO-Bench, Flora led on finding where actions begin and end. It trailed Astra overall and fell behind Gemini 3.8 Flash on pure description accuracy. On Nvidia's DenseCaptionBench, Flora came second to Astra for robot videos. The researchers flag a weakness in that test: Astra is both a contestant and the model that scores the descriptions.
The grader's mistakes land on someone
The most useful numbers are the error rates. Flora-Rewarder caught 53.3 percent of failed steps inside an episode. No competitor passed 11.7 percent, and Astra caught none. That is a real gain.
It also means nearly half of failed steps went uncaught. In addition, roughly 11 percent of steps that were actually normal got wrongly marked as errors. Each of those has a human cost. A missed failure reaches a deployed robot. A false flag sends an engineer to review work that was fine.
Descriptions can also be wrong. The Decoder gives one example: a robot hand wiping a plate with a cloth, which Flora describes as folding a napkin. The model can also chop a longer action into more small pieces than there should be. On a video with many short hand movements, though, it found 34 of 37 actions while Astra found four.
The lesson here is that cheap labeling moves the cost, it does not remove it. Spending falls on the labeling bill and rises on the work of checking the labeler. Whoever signs off on a robot deployment inherits that error profile.
What leaders and teams should ask
Ask whether any quoted labeling cost is a measured figure or a projection, and at what GPU price. Ask for the miss rate and the false-alarm rate on your own footage, not only on a vendor benchmark.
Ask who judged the benchmark. A model that grades its own competition gives a weaker result than one scored by humans. Ask which label type matters most for your use. Flora's reported edge is in failure detection and action timing. Astra still led on rubric-based scoring and trajectory ranking.
Finally, decide where a human stays in the loop. A grader that misses close to half of failed steps is a screening tool, not a safety case.
Labels used to be the cost of teaching robots. Now they are also the evidence for trusting them.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Nvidia (GEAR research lab).
