FLORA: An Open Model Suite for Scaling Robotics Data Engines
Sicong Leng, Yunze Man, Zhiqi Li, Jing Wang, Yuhe Zhang, Ryan Punamiya, Siddharth Patki, Kaiyuan Zheng and 9 others
NVIDIA
Published October 22, 2026 · on the feed October 26, 2026 · ▲ 2 · arXiv 2610.31779
The abstract
The life-cycle of robotics foundation models crucially depends on data labels. Outcome reward labels are needed for monitoring real-world deployments. Process reward labels are needed during RL post-training. Timestamped captions are needed for filtering and balancing pretraining data. We present Flora, an industrial-quality data engine used for training GR00T models, that simultaneously covers these capabilities. Flora achieves SOTA scores on the captioning benchmark WGO-Bench and reward benchmarks RoboReward and Robometer. On internal benchmarks it is competitive with GPT-6 Astra while beating other frontier models. The model is substantially cheaper than API-based alternatives, roughly 125 times cheaper than GPT-6 Astra. Florais battle tested at scale, labeling 200,000 hours of data per day across 512 GB300s. By open-sourcing our data engine, we enable researchers to scale custom robotics datasets and explore emerging applications such as physical auto-research and human-to-robot transfer.
The Brief
1:52A short podcast summary of the paper. Written and voiced by AI.
Transcript
FLORA. An Open Model Suite for Scaling Robotics Data Engines. By Sicong Leng and sixteen others, from NVIDIA. Published on the twenty-second of October. Its code repository is worked out: about eighty source files, with a readme and a licence.
Robotics foundation models depend on data labels at every stage of their life-cycle, and three kinds are named here. Outcome reward labels are needed to monitor deployments in the real world. Process reward labels are needed during reinforcement learning post-training. Timestamped captions are needed to filter and balance pretraining data. Flora is an open model suite that covers all three at once: an industrial-quality data engine, the same one used for training GR00T models, now open-sourced.
As a labeler, it is measured against public benchmarks and internal ones. Flora achieves state-of-the-art scores on the captioning benchmark WGO-Bench, and on two reward benchmarks, RoboReward and Robometer. On internal benchmarks it is competitive with GPT-6 Astra, while beating other frontier models. It is also substantially cheaper than API-based alternatives: roughly one hundred and twenty-five times cheaper than GPT-6 Astra.
Its provenance is production use. Flora is described as battle tested at scale, labeling two hundred thousand hours of data per day across five hundred and twelve GB300s. That is the volume at which the engine already runs while GR00T models are trained on what it labels.
The release is aimed at researchers who want to scale custom robotics datasets of their own. The authors also point to emerging applications it opens up, and name two: physical auto-research, and human-to-robot transfer.
Flora's case fits in one line: two hundred thousand hours of robot data labeled every day, at about one hundred and twenty-five times less than the cost of GPT-6 Astra. Opening that engine means outcome rewards, process rewards and timestamped captions can all come from one model that researchers run on their own datasets, rather than from a paid API.
More with an account
Listen to every day’s papers
Daily playlists, deep dives, quests for the papers, and quizzes on their key terms.
Join free