NVIDIA open-sources Flora, the robotics data engine behind GR00T: three models that caption and score robot video at roughly 1/125 the cost of GPT-6 Astra

Background and the problem

Robot foundation models need huge amounts of labeled data: timestamped action captions to filter and balance pretraining data, dense process rewards for RL post-training, and outcome rewards to monitor real deployments. Those labels have mostly come from humans or expensive frontier multimodal APIs, which does not scale to hundreds of thousands of hours of video. NVIDIA's GEAR lab is now releasing Flora, the data engine it uses to train GR00T, aimed squarely at that bottleneck.

Architecture and how it works

Flora is a three-stage pipeline. Flora-TempSeg 9B, a LoRA adapter on Qwen3.5-9B, localizes operations window by window and stitches them into episode-level events. Flora-Captioner 27B, built on Qwen3.5-27B with SFT followed by RL, writes one English action instruction per segment using surrounding context. Flora-Rewarder 27B outputs an episode success score, short-step action-quality scores and failure reasons. It handles both human egocentric and robot video, and NVIDIA says it labels 200,000 hours of video a day on 512 GB300s.

Benchmarks

All figures are NVIDIA's: on a 50-episode internal eval, Flora scores 57.1% quality versus 53.7% for GPT-6 Astra at maximum effort. On DenseCaptionBench's human subset, segmentation [email protected] is 54.6 versus 44.2 for Astra, though Astra leads on the robot subset at 71.9 versus 66.9. RoboReward MAE is 40.0 (lower is better) versus 55.4, and DenseRewardBench failure recall is 53.3% versus 0 for Astra. At $10 per GPU-hour on eight GB300s, Flora costs about $0.72 per hour of source video against about $88.49 for Astra, roughly 125x; throughput figures are modeled, not measured.

Getting started

All three weights are in the nvidia/flora collection on Hugging Face and run with transformers (TempSeg loads as a PEFT LoRA); NVIDIA asks users to keep the FP32 Gated DeltaNet parameters alongside BF16 base weights. They suit research on robot data cleaning, sub-task segmentation, RL reward signals and deployment monitoring. The license is NVIDIA OneWay Noncommercial, research only, and NVIDIA stresses the outputs are annotations, not robot control or a sole safety monitor. The paper PDF and full comparison tables are on the project page.