AI Tech Flash ── 2026-10-26
NVIDIA Flora: Robot Video Labels at $0.72 per Hour
NVIDIA's open Flora tool is estimated to label robot video for $0.72 per hour
What changes
According to NVIDIA, Flora could sharply cut the cost of annotating robot video. Annotation means writing down what happens in a video so a system can learn from it.
According to shattered.io, NVIDIA's repository estimates $0.72 per hour of video. GPT-6 Astra, the comparison model, is put at about $88.50, a gap of roughly 125 times. These are projections, not measured costs.
Why it matters
Robots learn from many demonstration videos, each broken into steps, described in words, and marked as success or failure. According to shattered.io, collecting footage is getting easier, while labeling is the slow part.
According to eGamers.io, large language models such as GPT-6 Astra can produce several kinds of labels but cost too much at robotics volumes. Cheaper labels would widen the range of projects that are affordable.
What it is
According to shattered.io, Flora is an open-source model suite and data engine published by NVIDIA on GitHub. A data engine is a system that processes data into a form fit for training.
It takes videos of robots and humans handling objects. It outputs action segments, captions, and an assessment of outcomes and rewards. A reward is a score showing whether an action went well.
There are three components. Flora-TempSeg splits video into segments, Flora-Captioner describes each segment, and Flora-Rewarder judges whether the action worked.
Who and where
NVIDIA built Flora. According to eGamers.io, the models are available on Hugging Face. Shattered.io reports that teams can run the suite on their own hardware.
According to eGamers.io, NVIDIA already uses Flora in training its GR00T models. NVIDIA said it labels up to 200,000 hours of video a day on 512 GB300 GPUs.
Why it was built
According to eGamers.io, NVIDIA's GEAR research lab sorts robot-video labels into three types. The first records which action happens at what time. The second scores progress step by step. The third decides whether a task succeeded.
The article says most current tools cover only one of the three. The Flora paper says it handles all three. The article also reports that NVIDIA presents Flora as an open alternative to paid services.
How it works
According to eGamers.io, Flora is three fine-tuned Qwen3.5 models. Qwen is a family of openly available language models.
A 9B model splits each video into action segments. B means billions of parameters, a measure of model size. A 27B model then describes every segment, using context from before and after it.
The third part, Flora-Rewarder, is another 27B model. Depending on the prompt, it rates an episode from 1 to 5, labels each half-second step as normal, suboptimal or failure, or scores against user-defined rubrics.
Key points
- According to shattered.io, NVIDIA's estimate is $0.72 per video hour, against about $88.50 for GPT-6 Astra.
- The estimate assumes eight GB300 GPUs rented at $10 per GPU-hour. It is not a measured cost.
- In a 50-video test reported as NVIDIA's, annotation quality was 57.1% versus 53.7%. No independent check exists.
Conclusion
According to NVIDIA's claims as reported, Flora aims for quality near that of a frontier model at a far lower cost. But both cost and quality figures come from NVIDIA's side, and no third party has confirmed them.
According to eGamers.io, Flora errs on long actions, sometimes splitting them or describing them wrongly. Teams should test it on their own footage.
Quotes and sources
“NVIDIA’s open-source Flora suite can label an hour of robot video for an estimated $0.72”── shattered.io, 2026-10-26
“Nobody measured those costs in production. Nvidia projected them, using an assumed rental rate of $10 per GPU per hour.”── eGamers.io, 2026-10-26
“Flora-Rewarder catches 53.3 percent of failed steps inside an episode.”── eGamers.io, 2026-10-26
“The researchers admit a flaw in the method: GPT-6 Astra grades the descriptions and is also one of the models under test.”── eGamers.io, 2026-10-26
Source articles
- shattered.io 10/26 09:55
- eGamers.io 10/26 07:58
Written by AI (Claude) from the cited articles, then checked by Jev (TypeSafe's judgement model), a Gemini fact-check and a verbatim check of every quote.