By Thorsten Meyer | October 25, 2026 | ThorstenMeyerAI.com
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Robots learn from video, but someone has to describe every second of it: what the hand did, when it did it, and whether it worked. Nvidia’s GEAR robotics lab has released Flora, three models that do this job, and says they do it for 72 cents per hour of video instead of $88 with GPT-6 Astra. I read the paper, the model cards and the license. The cost figure is a modeled estimate, the quality picture is mixed, and “open” here means free for research, not for business.
Key takeaways
- What it is: three video models built on Alibaba’s open Qwen3.5: a 9B segmenter that finds each action, a 27B captioner that describes it, and a 27B rewarder that scores how well the robot did. Nvidia says Flora labels data for its GR00T robot models, up to 200,000 hours of video per day on 512 GB300 GPUs.
- The 125× is Nvidia’s own estimate. Flora’s $0.72 per video hour is modeled on rented GPUs at an assumed $10 an hour; GPT-6 Astra’s $88.49 is measured API usage on the same 50 test episodes, about 75 minutes of video. The exact ratio is 122.5.
- Quality is competitive, not dominant. Flora wins on human videos, failed steps and partial success. GPT-6 Astra still wins on robot videos, caption accuracy and rubric scoring. Nvidia ran most tests itself.
- The license is noncommercial. All three models ship under the NVIDIA OneWay Noncommercial License: “research or evaluation purposes only,” and derivatives inherit that limit. Nvidia itself may use them commercially. For companies, Flora is a yardstick and a blueprint, not a free production pipeline.
- For users: cheaper labels remove one bottleneck on the way to household and service robots. Collecting robot data, safety and hardware cost remain hard.
Products Worth Considering
Amazon Kindle Paperwhite (2024), 16GB, 7-inch Display, Black
LG UltraGear 27G640A-B 27-Inch QHD IPS Gaming Monitor, 300Hz, 1ms, G-SYNC Compatible, FreeSync Premium, DisplayHDR 400
Sonos One (Gen 2) Two-Room Set Voice-Controlled Smart Speaker with Amazon Alexa Built-in (2-Pack, Black)
As an affiliate, we earn on qualifying purchases.
The job nobody sees
Robot foundation models learn from recordings: a gripper placing a mug, two robot hands passing a wire, a person folding a towel, filmed from their own eyes. According to the Flora paper, training and running these models needs three kinds of labels:
- Timestamped captions, such as “Place the blue mug on the tray with your right hand,” tied to the seconds where it happens. They let engineers filter and rebalance huge training sets.
- Process rewards, a grade for each short step: normal, suboptimal or failure. Reinforcement learning uses them to find the moment a robot went wrong.
- Outcome rewards, a score for the whole attempt, used to monitor robots after deployment.
Most tools deliver one of these. Frontier models such as GPT-6 Astra can deliver several, but Nvidia says they are “expensive at the scale of modern robot datasets.”
What Flora is
Flora is three models on Hugging Face, with a project page and paper from the GEAR lab of Yuke Zhu and Linxi “Jim” Fan.
- Flora-TempSeg (9 billion parameters, an adapter on Qwen3.5-9B) returns time intervals: where one operation ends and the next begins.
- Flora-Captioner (27 billion parameters, fine-tuned from Qwen3.5-27B) writes one English instruction per interval, using the video before and after for context. It takes one camera, or a main camera plus one on each wrist.
- Flora-Rewarder (27 billion parameters) scores an episode from 1 to 5, grades every half-second as normal, suboptimal or failure, or scores against a user’s rubric.
Per the model cards, Flora was the data engine for GR00T-2 and was trained on under 10,000 hours of robot and human video. The inference code ships inside the Hugging Face repositories; there is no separate GitHub repo. Test hardware is an Nvidia H100 on Linux. By my estimate, each 27B model needs roughly 55 GB for its weights in 16-bit precision: data-center territory, not a gaming PC.
Infographic 01 · What Flora does
Three models that turn robot video into training labels.
Flora cuts a recording into actions, describes each one and grades how well it went. The 200,000-hour figure is Nvidia’s own production claim.
Flora-TempSeg
9B · adapter on Qwen3.5-9B
- Job: Finds where each action starts and ends
- In: Video windows (one camera, or main + two wrist cameras)
- Out: Time intervals as JSON
Flora-Captioner
27B · fine-tuned Qwen3.5-27B
- Job: Describes each action, using the video before and after
- In: One interval plus context
- Out: One English instruction, e.g. “Place the blue mug on the tray with your right hand.”
Flora-Rewarder
27B · fine-tuned Qwen3.5-27B
- Job: Judges how well the robot did
- In: An episode or a 16-second window, plus the task
- Out: Score 1–5 per episode · normal / suboptimal / failure per 0.5 s · rubric scores
Three labels, three uses
Why robot models need each one, per the paper
| Label | Used for | Made by |
|---|---|---|
| Timestamped captions | Filtering and rebalancing pretraining data | TempSeg + Captioner |
| Process rewards | Reinforcement learning after pretraining: which step went wrong | Rewarder |
| Outcome rewards | Monitoring robots after deployment | Rewarder |
What it is not
model cards
- Not a robot controller: outputs are annotations only
- Not meant to be the sole safety monitor for a physical system
- Not for commercial or production use (noncommercial license)
- Hardware: tested on Nvidia H100, Linux; ~55 GB of weights per 27B model (estimate)
Sources: Flora paper, Nvidia (20 Oct 2026); Hugging Face model cards (15 Oct 2026); Nvidia GEAR project page. Weight-size estimate: Thorsten Meyer (27B parameters × 2 bytes).
Powered by Thorsten Meyer AI
How the 125× was calculated
Nvidia deserves credit for documenting this in Appendix D of the paper.
- The test set: 50 fixed episodes from Nvidia’s own DenseCaptionBench, 20 with human hands, 15 with two-finger grippers and 15 with dexterous robot hands. Total: 4,483 seconds, about 75 minutes of video.
- The baseline: GPT-6 Astra at maximum reasoning effort, at public API rates from October 4, 2026: $88.49 per hour of video. 106 of 7,062 attempts reported no usage, so this is a known-usage subtotal.
- Flora’s cost: modeled, not measured: eight GB300 GPUs at an assumed $10 per GPU-hour and a modeled throughput. Result: $0.72 per video hour, with scenarios from $0.51 to $1.26, excluding storage and overhead. In the paper’s words: “The production GPU-hour denominator has not been verified.”
- The ratio: $88.49 divided by $0.72 is 122.5, rounded to “approximately 125×.”
- Quality on the same episodes: Flora 57.1%, Astra 53.7%, Gemini 3.8 Flash 38.3% (at $8.48 per hour), Claude Opus 5 30.4% (at $41.70).
The gap is real: even the pessimistic scenario is about 70 times cheaper than Astra, and Flora is about 12 times cheaper than Gemini 3.8 Flash while scoring higher. But the headline compares a measured API bill with a modeled GPU budget, on 75 minutes of video the vendor chose and scored. And it covers captioning only; the reward model is not part of the 125×. The Decoder also calls the figures “projections, not production measurements.”
Where Flora wins, and where it doesn’t
The paper’s full benchmark table tells a more mixed story. Flora leads on:
- Action boundaries and human videos. Segmentation on the public WGO-Bench: 50.3 vs Astra’s 46.0. Human subset of Nvidia’s DenseCaptionBench: 37.4 vs 29.7.
- Partial success. On the public RoboRewardBench, Flora’s error is 27.8% lower than Astra’s. Example: a robot topples a bottle but fails to knock it off the table. Flora gives the human reference score of 3 out of 5; Astra gives 1.
- Catching failures. On Nvidia’s internal DenseRewardBench, Flora finds 53.3% of failed steps. No other model tested exceeds 11.7%, and Astra finds none. The price: Flora also flags about 11% of normal steps as failures.
GPT-6 Astra leads on:
- Robot videos. Robot subset of DenseCaptionBench, all three metrics; joint score 57.9 vs 51.0.
- Caption accuracy. On WGO-Bench, given the correct time window: 77.4% vs 64.7%. Even the budget model Gemini 3.8 Flash scores 70.9%.
- Ranking and rubrics. Ranking unfamiliar robot trajectories (82.3 vs 72.0) and matching human rubric grades (error 3.6 vs 6.5).
Two caveats. First, on the caption benchmarks a GPT-6 Astra judge grades the answers while Astra also competes; the researchers name this “generator–judge self-preference” as a limitation. If anything it favors Astra, but none of the caption scoring is independent. Second, the examples cut both ways. In one selected episode, Flora found 34 of 37 reference actions where Astra found four. In another, a robot hand wipes a plate with a cloth, and Flora calls it folding napkins.
Infographic 02 · The 125× claim, unpacked
A measured API bill against a modeled GPU budget.
How Nvidia computed “125× cheaper than GPT-6 Astra,” and where Flora wins and loses in its own paper.
Cost per hour of video
modeled estimate vs measured API bill
Flora scenarios: $0.51–$1.26. Storage and overhead excluded. Paper: “The production GPU-hour denominator has not been verified.” Covers segmenting and captioning, not the reward model.
Quality on the same 50 episodes
Equal mix of segmentation, captions and joint score
Caption scores are judged by GPT-6 Astra, which also competes. The researchers name this self-preference as a limitation.
Who wins where
Selected rows from the paper’s Table 1. Higher is better unless noted.
| Benchmark | Flora | Astra | Ahead |
|---|---|---|---|
| Segmentation, WGO-Bench public | Flora 50.3 | Astra 46.0 | Flora |
| Human videos, DenseCaptionBench Nvidia | Flora 37.4 | Astra 29.7 | Flora |
| Outcome error, RoboRewardBench public, lower is better | Flora 40.0 | Astra 55.4 | Flora |
| Failed steps caught, DenseRewardBench Nvidia | Flora 53.3% | Astra 0.0% | Flora |
| Robot videos, DenseCaptionBench Nvidia | Flora 51.0 | Astra 57.9 | Astra |
| Caption accuracy, WGO-Bench public | Flora 64.7 | Astra 77.4 | Astra |
| Rubric error, RubricBench Nvidia, lower is better | Flora 6.5 | Astra 3.6 | Astra |
Trade-off: Flora also flags about 11% of normal steps as failures.
Real gap, modeled number. Even the pessimistic scenario is about 70× cheaper than Astra.Assessment: Thorsten Meyer
Source: Flora paper, Nvidia (20 Oct 2026), Figure 3, Section 2.3, Table 1 and Appendix D. API prices: public standard rates as of 4 Oct 2026, per the paper.
Powered by Thorsten Meyer AI
The turn: data is the bottleneck, and Nvidia sells the shovels
Chatbots learned from text scraped off the web. Robots can’t: every hour of a robot hand folding a towel has to be recorded, then labeled. That has become a business. Scale AI sells a Physical AI Data Engine for collection and model-assisted annotation, and Nvidia’s paper names it as the commercial reference point.
Flora moves part of that work from people and API bills to GPUs you control. That also serves Nvidia, as The Decoder points out: free models that run on Nvidia GPUs help sell Nvidia GPUs. It fits a pattern, from Nvidia’s Nemotron 3.5 open models to its SimReady library of simulation assets for robotics. And as with Meta’s Muse Spark Open, the word “open” needs a look at the license.
What this means for companies
This section is my assessment.
- Read the license before the benchmark. The NVIDIA OneWay Noncommercial License says the models “only may be used or intended for use non-commercially,” defined as “research or evaluation purposes only.” Derivatives, including your fine-tunes, carry the same limit. Labeling your product’s training data with Flora is off the table unless Nvidia agrees otherwise; the model cards offer no commercial path.
- Use it as a yardstick. Evaluation is allowed. Run Flora on a sample of your footage and compare it with your current vendor or API pipeline.
- Build vs. buy just got a blueprint. The paper documents the recipe: segment, caption with context, then reward. The base models, Qwen3.5-9B and 27B, are under Apache 2.0. A company with its own labeled footage can follow that recipe on a permissive base. The weights are restricted; the method is published.
- Redo the cost math with your numbers. Nvidia’s $0.72 assumes GB300s at $10 an hour and excludes storage and overhead. Use your real GPU price, utilization and engineering time, and budget human review: an 11% false-alarm rate means people still check.
- Manufacturers and logistics operators already record hours of robot-cell footage; automatic outcome scoring could flag failed picks. But the model cards say Flora is not meant to “serve as the sole safety monitor for a physical system,” a caution that applies to any tool like it.
- Robotics startups: your moat moves. If labeling gets cheap, the advantage shifts to whoever owns the footage and the robots to record more.
Infographic 03 · Before you build on it
Free to study. Not free to ship.
What the license allows, what companies can do with Flora today, and what it changes for people waiting for useful robots.
The license
noncommercial NVIDIA OneWay, 22 Mar 2022
- Allowed: “research or evaluation purposes only”
- Not allowed: commercial use, production pipelines
- Derivatives such as fine-tunes inherit the limit
- Nvidia itself may use it commercially
- Base models Qwen3.5-9B / 27B: Apache 2.0
For companies
assessment build vs. buy
- Benchmark your vendor or API pipeline against Flora on your own footage (evaluation is allowed)
- Copy the recipe, not the weights: segment, caption with context, reward, on a permissive base
- Redo the math with your GPU price, utilization and staff time
- Keep humans in review: about 11% false alarms on failure steps
- Never the sole safety monitor on a robot cell
For users
closer, not here
- One bottleneck down: labeling. Data collection, safety and hardware cost remain
- No kitchen robot next year because of this, but faster training loops
- Your hands may be training data: ask what a work or app camera records and trains
- Students and researchers: free weights and a public paper, for noncommercial work
“Open” here means open weights for research. Read the license before the benchmark.Assessment: Thorsten Meyer
Sources: NVIDIA OneWay Noncommercial License; Hugging Face model cards (15 Oct 2026). Recommendations: Thorsten Meyer.
Powered by Thorsten Meyer AI
What this means for users
- Closer, not here. Household and service robots need enormous amounts of labeled video. Cheaper labeling removes one bottleneck, not the others: recording robot data is slow and costly, robots must be safe around people, and hardware must get affordable. Flora won’t put a robot in your kitchen next year. It speeds up the training loop that might.
- Your hands may be training data. Flora also labels video filmed from people’s own perspective, for research on transferring human skills to robots. The model cards ask for “proper rights and permissions” for any input video. As labeling gets cheap, everyday footage of people at work gains value. If a job or an app asks you to wear a camera, ask what the footage will train.
- Students and researchers get free weights and a documented paper to build on, as long as the work stays noncommercial.
Honest limits
- Confirmed from primary sources: model sizes, inputs, outputs, license, benchmark numbers and the cost method.
- Vendor claims, not independently verified: the 125× figure, the 200,000 hours per day on 512 GB300 GPUs, and Flora’s role in GR00T training. As of October 25, I found no independent benchmark. Humanoids Daily also flagged the noncommercial terms but reports Nvidia’s numbers without testing them.
- Whose benchmarks: DenseCaptionBench, DenseRewardBench and RubricBench are Nvidia’s own; WGO-Bench, RoboRewardBench and RobometerBench are public.
- Not tested by me: I did not run the models; the hardware estimate is my calculation.
Verdict
Flora is a useful release with an unusually candid paper. The 125× is a modeled estimate on 75 minutes of video, and the real gap is still large. The most interesting result isn’t the price; it’s a reward model that catches failed steps the frontier models miss. But “open-sourcing” is the wrong word for a noncommercial license. Researchers get a gift. Companies get a benchmark, a recipe and a reminder: in robotics, whoever owns the footage owns the future, and Nvidia would like that footage processed on Nvidia GPUs.
Sources
- Nvidia GEAR: Flora: An Open Model Suite for Scaling Robotics Data Engines, project page (accessed October 25, 2026)
- Leng, Man, Li et al. (Nvidia): Flora paper, PDF (dated October 20, 2026; cost method in Appendix D)
- Nvidia on Hugging Face: Flora-Captioner, Flora-Rewarder and Flora-TempSeg model cards (release date October 15, 2026, per the cards)
- Nvidia: NVIDIA OneWay Noncommercial License (22 March 2022)
- Macrodata Labs: WGO-Bench (public benchmark used in the paper)
- The Decoder: Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra (October 25, 2026)
- Humanoids Daily: NVIDIA Releases Flora to Turn Robot Videos Into Training Data (October 21, 2026)
- Scale AI: Expanding Our Data Engine for Physical AI
