Nvidia's Flora labels robot video 125x cheaper, by its own estimate

What this is about

Nvidia has released Flora, an open data engine for robotics. It captions and scores the videos used to train large AI models for robots. It was developed by Nvidia's GEAR research lab, and the models are available on Hugging Face. Nvidia claims an hour of robot video costs about 72 cents, while OpenAI's GPT-6 Astra charges around $88 for the same work. That would make it up to 125 times cheaper. The comparison is more nuanced than the price suggests, though.

Why robot videos need labels

Robot models learn from videos that have to be annotated with text. According to GEAR, that takes three kinds of information:

• Captions describing which action happens when in the video. They are used to select training data.
• Scores for how well a robot progresses, step by step. With them, it can keep learning through trial and error (reinforcement learning).
• Verdicts on whether a task ultimately succeeded. They are used to monitor robots in deployment.

Existing tools usually deliver only one of these. Large language models such as GPT-6 Astra can produce several, but they are too expensive for the huge volumes of video involved. Flora is meant to cover all three.

How Flora is built

Flora consists of three fine-tuned models based on Qwen3.5:

• A model with 9 billion parameters (parameters are, roughly speaking, the knobs a model tunes during training) splits videos into action segments.
• A 27B model describes those segments, taking into account what happens before and after.
• Another 27B model, Flora-Rewarder, takes on three roles depending on the prompt. It rates an episode's success on a scale of 1 to 5, classifies each half-second step as normal, suboptimal or error, and awards points against rubrics you define yourself.

According to Nvidia, the system is already used to train the GR00T models and annotates up to 200,000 hours of video per day on 512 GB300 GPUs.

Costs and comparison

In a test with 50 videos, Nvidia says Flora captioned slightly better than GPT-6 Astra, even though Astra was allowed to use its full reasoning effort.

Important: the Flora cost is a projection. Nvidia didn't measure it in production. It calculated it using an assumed rental price of $10 per GPU-hour.

Strengths and weaknesses in the benchmarks

Video captions: On the public WGO-Bench, Flora is best at detecting action boundaries. Overall it trails Astra, and on pure caption accuracy it also trails Gemini 3.8 Flash. The researchers say the gains come mainly from temporal localization. On Nvidia's own DenseCaptionBench, Flora leads on every metric for human footage and comes second behind Astra on robot videos. The researchers admit one catch: the captions are judged by GPT-6 Astra, which is itself part of the test.

Scoring: On RoboRewardBench, Flora's error is 27.8 percent lower than Astra's. The advantage comes from partially completed tasks: if a robot only tips a bottle over instead of pushing it off the table, Flora gives 3 out of 5, like the human reference, while Astra gives just 1.

Error detection: The gap is widest for errors within an episode. Flora-Rewarder finds 53.3 percent of error steps, no competitor gets above 11.7 percent, and Astra doesn't catch a single one. On the downside, Flora wrongly flags about 11 percent of normal steps as errors. Astra leads on rubric scoring and on ranking trajectories.

Examples: In a video with many short hand movements, Flora detects 34 of 37 actions, while Astra finds only four. On the other hand, Flora sometimes splits longer actions into too many small steps and occasionally gets things wrong: a robot hand wipes a plate with a cloth, and Flora describes it as folding a napkin.

What you need to try it

Nvidia doesn't list hardware requirements for individual users. As a rough guide: the 27B model needs about 17–20 GB of VRAM with 4-bit quantization, and the 9B model about 7 GB. For large-scale operation, Nvidia uses 512 GB300 GPUs. The source doesn't say whether quantized variants or GGUF files exist.

Our take

Flora is mainly interesting for robotics researchers who want to scale their own datasets, for example to transfer human demonstrations to robots. Nvidia positions it as an open alternative to commercial services such as Scale AI's Physical AI Data Engine. There is a business interest behind it: Nvidia is investing billions in open models to keep developers in its own hardware ecosystem and sustain chip demand. Flora adds detailed captions and reward signals to the data pipeline shown at GTC in the spring, and fits Nvidia's strategy of covering the entire robotics value chain. Google's recently unveiled Gemini 4 Argon addresses something similar.

The caveats: all the numbers come from Nvidia, the costs are estimates, the test covered only 50 videos, and on DenseCaptionBench a competing model acts as judge. The 125x price advantage also holds only under Nvidia's assumptions. The strong error detection is a real plus, but it comes at the price of false alarms.