NVIDIA open-sources Flora: robot data labeling 125 times cheaper than GPT-6 Astra

NVIDIA has open-sourced Flora, the data engine it uses to train its GR00T robotics foundation models. The system labels 200 000 hours of data a day on a cluster of 512 GB300 GPUs. According to the authors, it is about 125 times cheaper than the GPT-6 Astra API and matches it on internal benchmarks. Flora outperforms the other frontier models. On the public benchmarks WGO-Bench, RoboReward and Robometer, it reaches SOTA.

Three kinds of labeling in one model

A robotics model depends on labels at every stage, and each stage needs a different kind. Pretraining requires captions with timecodes. They are used to filter and balance the video corpus so it doesn't contain too many near-identical clips of "robot moves a cube." Monitoring real-world deployments requires outcome reward labels: did the robot complete the task or not. RL fine-tuning requires process reward labels, meaning an assessment of whether each step brings the robot closer to the goal.

These tasks are usually handled separately. Flora covers all three with a single model, so the same pipeline can plug into pretraining, production evaluation and reinforcement-learning fine-tuning.

The arithmetic of scale

200 000 hours a day on 512 GB300s works out to roughly 390 hours of video per GPU per day, or about 16 hours of footage for every hour of GPU time. These are my calculations from the stated figures. They explain where the price gap comes from. Running hundreds of thousands of hours of video a day through a closed API makes no economic sense, while a specialized model on your own hardware does it for a fraction of the cost.

The key comparison here is with GPT-6 Astra. Quality parity at a price 125 times lower means labeling stops being a line item that caps dataset size.

What exactly is claimed

• Flora: an open set of models and a production-grade data engine on which the GR00T models are trained.
• Three functions: captions with timecodes, outcome reward labels, process reward labels.
• WGO-Bench captioning benchmark: SOTA.
• RoboReward and Robometer reward benchmarks: SOTA.
• Internal tests: on par with GPT-6 Astra, ahead of the other frontier models.
• Cost: about 125 times lower than GPT-6 Astra.
• Throughput: 200 000 hours of data a day on 512 GB300s.
• Release: code in the NVIDIA/Flora repository on GitHub, project page at research.nvidia.com/labs/gear/flora.
• Authors: Sicong Leng, Yunze Man, Zhiqi Li, Jing Wang, Yuhe Zhang, Ryan Punamiya, Siddharth Patki, Kaiyuan Zheng, Zi-ang Cao, Zhiding Yu, Sifei Liu, Danfei Xu, Ben Levin, Jan Kautz, Yuke Zhu, Johan Bjorck and Linxi "Jim" Fan.

Who can use it right now

The biggest winners are teams that have their own robot recordings but no budget to label them with closed models. Academic labs and startups can build their own dataset, filter it by timecodes and get a reward signal for RL without external APIs. Filtering and balancing pretraining data is no longer something only the big players can do.

The authors name two directions this opens up: physical auto-research and skill transfer from humans to robots. In the first, reward labeling lets a robot evaluate the results of its own experiments. In the second, the model labels videos of people so that they can be used to train manipulators.

For practitioners, the test is simple: take the repository, run it on your own recordings, and compare the quality of captions and rewards with what an API currently does for a fee. If the claimed parity with GPT-6 Astra holds up on other people's data, NVIDIA's bet on an open data engine around GR00T will be very convenient for the whole industry. The same 125 times price gap, but now in your own calculations.