Nvidia’s Flora Labels Robot Video 125× Cheaper Than Astra

Nvidia’s GEAR robotics lab has released Flora, a set of three models that turn raw demonstration video into the labels robot models are trained on. The paper, dated 20 October, describes it as the data engine Nvidia already uses to train its GR00T models, labelling up to 200,000 hours of video a day across 512 GB300 GPUs. The weights are on Hugging Face under Nvidia’s own model licence, and the authors include Linxi “Jim” Fan and Yuke Zhu.

Flora produces three kinds of label that robotics teams usually pay humans or expensive API models for. Timestamped captions describe each action in a clip and are used to filter and balance pretraining data. Process rewards grade an episode step by step during reinforcement-learning post-training: every half-second step is marked normal, suboptimal or failure. Outcome rewards judge whether the task succeeded, which is what you need to monitor robots once they are deployed. Each job has its own model, all fine-tuned from Alibaba’s Qwen3.5: a 9B model that cuts video into action segments, a 27B captioner and a 27B Flora-Rewarder.

The headline number is cost. On a fixed set of 50 episodes, Nvidia scores Flora at 57.1% annotation quality for an estimated $0.72 per hour of source video, against 53.7% for GPT-6 Astra at maximum reasoning effort, which cost $88.49 per hour at public API prices. Claude Opus 5 reached 30.4% for $41.70 an hour and Gemini 3.8 Flash 38.3% for $8.48. The paper is careful about what that comparison is: the API figures are measured usage, while Flora’s is a projection that assumes an eight-GPU GB300 deployment rented at $10 per GPU-hour, with a range of $0.51 to $1.26 depending on the scenario. The exact ratio to Astra is 122 to 1, rounded to “125×” in Nvidia’s summary.

Quality is more mixed than the price gap. Flora beats every model tested on human video and on step-level rewards, and its most striking result is catching failures: Flora-Rewarder flagged 53.3% of the steps where something went wrong, while Astra caught none and no other model passed 1%. That comes at a price, since it also wrongly flags about 11% of normal steps. On robot video Astra is still ahead, 57.9 to 51.0 on the joint caption score, and it wins on WGO-Bench and on rubric-based scoring, where its error is 3.6 against Flora’s 6.5. Nvidia also notes that Astra helped evaluate the captions it is being compared on.

Flora extends the physical-AI stack Nvidia has been opening up all year, from the Cosmos 3 model and GR00T humanoid in June to the Nemotron 3.5 agent models this month. The paper opens from the premise that a robot model’s whole life cycle, from pretraining to deployment, depends on labels, a point The Decoder also drew out in its write-up. If a 27B model can grade a robot’s every half-second for under a dollar an hour, the cost of turning recorded video into a training signal stops being the limit on how much of it gets used.