Grading robot video is where a physical-AI program starts to cost real money. These are robots trained from footage, not a chatbot. NVIDIA’s projected cost for Flora, its data engine for that job, is about 72 cents an hour, against roughly $88 to $88.50 for GPT-6 Astra. The Decoder, in Jonathan Kemper’s October 25, 2026 report, calls that up to 125 times cheaper, by NVIDIA’s own estimates. The printed prices do not yield 125. They divide out to about 122, roughly one 122nd of the reported API price. The Flora side is a projection, not a production bill.
Those prices, a 50-video comparison, a daily throughput claim, and the benchmark rankings below are NVIDIA’s results as told by The Decoder. They are not on the Flora-rewarder-27B card NVIDIA posted to Hugging Face on October 15, 2026. The card is the document that tells you what you may actually do with the weights.
The license decides the budget
The published weights sit under the NVIDIA OneWay Noncommercial License, dated March 22, 2022. The card limits them to research and development, and states the restriction in one line:
Flora is not intended or supported for commercial or production use.
Apache 2.0 covers the Qwen base. It does not automatically cover NVIDIA’s fine-tune. Flora-rewarder-27B is a full fine-tune of Qwen3.5-27B, vision encoder included, and it is not an Alibaba product. NVIDIA using Flora internally while training GR00T-2 does not give a customer permission to run a labeling line.
Three other limits belong in the same meeting. The model is English only. It does not drive a robot, and NVIDIA says not to use it as the only safety monitor on a physical system. A legal experiment still means a research run on NVIDIA GPUs. The card lists Ampere, Hopper, Lovelace, and Blackwell. That is not an API you hand a vendor.
Training labels mixed human annotation, automatic success signals, and labels written by other vision-language models, drawn from open datasets, vendor-licensed video, and NVIDIA’s own footage. No mix ratio is published. It is fair to ask how much of a cheap labeler was taught by expensive ones. That remains an open question. It is not a finding.
What you can download, and what you can only read about
Downloadable and documented today is the rewarder: 27 billion parameters, text scores from RGB video plus a prompt. The prompt picks the format.
An outcome score is an integer from 1 (complete failure) to 5 (complete success), optionally with a rationale. A dense score, also called a process score, looks at a 16-second window and marks quality every half second: 1 for failure, 2 for suboptimal, 3 for normal, plus the worst score in the window. A rubric score returns a reasoning trace and one score per rubric step. The card does not say the model invents those rubrics. The prompt sets the job.
Dense scores are quality marks on fixed time steps. That is different from finding where an action starts and ends, or naming the action. Kemper’s report, citing NVIDIA’s GEAR lab, splits robot-learning labels into three jobs: timed action descriptions for choosing training clips, step-by-step progress ratings for trial-and-error learning, and success judgments for watching robots after deployment. Most tools, in that account, cover one. The rewarder card covers the second and third, as scores, not as timed captions.
The Decoder describes a wider Flora: a 9-billion-parameter Qwen3.5 model that segments actions, a 27-billion-parameter model that writes descriptions from the moments before and after an action, and Flora-Rewarder as the scoring model. Those two companions appear in the coverage. This card does not document them.
NVIDIA calls Flora the data engine used to train GR00T-2, and says the release is meant to help other researchers scale their own robot datasets. If a slide only says “GR00T, ” ask which one. An earlier GR00T N1 white paper is a separate document. Human-to-robot transfer is listed as an intended research use. The card does not report a result.
Why one hourly price is a blunt instrument
Sampling is not one setting. That is why a single “per video hour” number flattens three jobs. Outcome prompts can mosaic up to four camera views and take as many as 64 frames at about 10 frames a second. Dense prompts use 64 frames at 4 frames a second across that 16-second window. Rubric prompts treat each camera as its own video, at 1 frame a second. Any serious cost check needs video length, camera count, frame rate, windowing, and batching. The report does not supply them.
The Decoder says Flora’s costs are projections, calculated from an assumed GPU rental price of $10 an hour, not measured production bills. A caption in the same report puts the pair at $0.72 and $88.49, a few cents off the $88 to $88.50 range in the body. “Up to 125 times cheaper” is NVIDIA’s headline estimate, not arithmetic the write-up shows. “Up to” leaves room for a different configuration. It does not leave a purchase order.
The card publishes only upper bounds on training. It does not give a size. Fewer than 10, 000 hours of video. Fewer than a million examples. Fewer than a billion text tokens. The corpus could sit near those caps or far below them. Training scale is unknown.
The same report says Flora is already annotating as many as 200, 000 hours of video a day across 512 GB300 GPUs while GR00T models train. The card supports the GR00T-2 data-engine role. It does not confirm the daily volume, the resolution, or whether that figure is a peak or a sustained run. Do not mix the two scales. A six-figure daily labeling number, if it holds, is inference. It does not show that the model learned from that much footage.
Better at the clock, weaker at the name
The skill split in that report is not a clean win for Flora. Researchers there credit the gains mainly to temporal localization: spotting when an action starts and ends.
On a public benchmark the report calls WGO-Bench, Flora leads at finding action boundaries, trails Astra on the overall score, and trails Google’s Gemini 3.8 Flash on pure description accuracy. On NVIDIA’s DenseCaptionBench, Flora leads the reported metrics for human recordings and ranks second to Astra on robot video. In that setup, Astra is also an evaluator of the descriptions it is being scored against. That does not void the bench. A buyer should still want a human pass before treating the ranking as settled.
A caption figure of 57.1 percent annotation quality for Flora against 53.7 percent for Astra is a separate number. The report does not define it, and it does not tie the figure cleanly to the 50-video test. In that 50-video test, NVIDIA is said to have found Flora slightly ahead even when Astra used its full reasoning budget. Anthropic’s Claude Opus 5, in the same reported test, scored much worse at about $42 an hour. Slightly ahead on an undefined caption metric is a different claim from leading WGO-Bench. Flora does not lead that bench.
One clip is not a rate. In a single video full of short hand movements, Flora is said to have found 34 of 37 reference actions. Astra found four. The misses cut the other way. Flora can chop a long action into too many tiny steps. In one case, a robot hand wiping a plate with a cloth is described as folding a napkin. Astra still leads, in that account, on rubric scoring and on ranking trajectories, and it often names actions more precisely even when it misses their boundaries. A labeler that knows when something happened and a labeler that knows what to call it are not the same hire.
Keep the score scales apart. Outcome scores run from 1 to 5. Process scores run from 1 to 3. On a reward test The Decoder calls RoboRewardBench, Flora’s error rate is reported 27.8 percent lower than Astra’s. That is a relative cut, not 27.8 points. A bottle that is tipped, not fully knocked over, is scored on the 1-to-5 scale: Flora gives 3, matching the human reference, while Astra gives 1.
Mid-episode failure is the point of a process score. Grade only the last frame and a robot can botch the middle, recover badly, and still look fine in the log. Flora-Rewarder is reported to catch 53.3 percent of failed steps inside an episode. No competitor in that comparison exceeds 11.7 percent, and Astra catches none. Catching 53.3 percent still means nearly half the failed steps slip through. The reported test shows better detection. It does not show reliable detection. The same account says Flora falsely flags about 11 percent of normal steps as errors.
NVIDIA’s card names RoboReward and RoboMeter as evaluation references. It does not publish those percentages. Until another lab repeats the work, they remain NVIDIA’s results as told by one publication. The card’s operational line does not depend on them: do not use Flora as the only safety monitor. You want both the miss rate and the false-alarm rate before you trust either.
Questions worth asking before anyone moves a budget
- What is Flora, stripped of the headline?
NVIDIA’s robotics data engine for turning video into labels and reward scores, used in training GR00T-2. The release you can inspect today is Flora-rewarder-27B, which scores episode success from 1 to 5, half-second action quality from 1 to 3, and rubric steps. The Decoder describes two further models for segmentation and description.
- Is it really 125 times cheaper than GPT-6 Astra?
That “up to 125 times” figure is NVIDIA’s estimate, as reported by The Decoder. The cost is a projection from a $10 GPU-hour assumption, not a measured bill. The printed prices, about 72 cents versus roughly $88 to $88.50, come out near one 122nd of the reported API price, and they ignore how differently each scoring mode samples video.
- Does the cheaper option label better?
On NVIDIA’s tests as relayed by The Decoder, Flora is stronger at action boundaries and at catching mid-episode failures, while Astra tends to describe actions more precisely and leads on rubric scoring and trajectory ranking. The 57.1 percent figure is a separate, undefined caption metric, and that reporting does not include a third-party replication.
- Can the released model go into a production pipeline?
No. The published weights are noncommercial research only, English only, and NVIDIA says they are not for production use or for sole safety monitoring. A research GPU run is not a labeling contract, and NVIDIA’s use of Flora while training GR00T-2 does not extend that right to customers.
- Will cheaper labels automatically produce better robots?
Not on the evidence published so far. The card does not show how much Flora improved GR00T-2 downstream, and human-to-robot transfer is a research aim, not a reported win.
- What will the card not tell you?
The mix of human labels versus model-written labels, any measured gain on GR00T-2, and any result on transferring human demonstrations to robots. It also publishes training scale only as upper bounds, so you cannot tell whether the rewarder saw 10, 000 hours or a small fraction of that.
What to do with it on Monday
If a frontier API is narrating your robot video, the gap in that report is large enough to audit the invoice. It is not large enough to book as savings. The released model is English only. A legal run is a research job on an NVIDIA GPU. It is not a labeling contract.
Ask for three things before anyone moves money. Score a sample of your own episodes with a person, with the API you already pay, and with any research model you are allowed to run. Count missed failures against false alarms. Split the cost across outcome scores, half-second process scores, and rubric passes. One blended hour hides the sampling differences.
A bigger model is no longer the only scarce input. You also need supervision you can trust when the footage scales. Flora is a serious research attempt at that supervision, with a reported edge in timing and in-progress failure, and a license that keeps the published weights in the lab. Until the projection becomes a bill, another lab repeats the benchmarks, and the license says you may ship it, do not sign the PO.