Overview
Flora is a set of models published on October 25, 2026 by NVIDIA's GEAR lab that labels video of robot tasks for training. It does not drive robots. It produces annotations: when each action happened, how well each step went, and whether the attempt succeeded.
Robot datasets need three kinds of labels and most tooling covers one. Timestamped captions describe actions and are used to filter and rebalance pretraining data. Step-level process scores guide reinforcement learning by pointing at the step that went wrong. Episode-level outcome judgements monitor robots after deployment. Flora covers all three with three models sharing the work.
All three build on Qwen3.5. Flora-TempSeg is a 9B LoRA adapter that finds where actions start and end. Flora-Captioner, at 27B, writes timestamped descriptions with surrounding context. A second 27B model, Flora-Rewarder, switches between three roles by prompt: scoring an episode from 1 to 5, classifying every half second as normal, suboptimal or failure, and grading against a custom rubric.
The licence deserves its own paragraph. All three ship under the NVIDIA OneWay Noncommercial License, which permits research and evaluation, and derivatives inherit the same restriction. This is a research tool, not a free labeling service you can drop into a commercial pipeline.
Key Features
- Three label types in one suite: Timestamped action descriptions, step-level process scores and episode outcome judgements, matching data filtering, reinforcement learning and deployment monitoring.
- 9B segmenter finds action boundaries: Flora-TempSeg is a LoRA adapter on Qwen3.5-9B. It takes a video window from one camera or a main camera plus two wrist cameras and returns time intervals as JSON.
- 27B captioner writes timestamped descriptions: Flora-Captioner describes an action with what came before and after it, returning one instruction-style English sentence per segment.
- One rewarder, three roles: Flora-Rewarder switches on prompt: a 1 to 5 score per episode, a normal, suboptimal or failure label every half second, and rubric scoring against criteria you define.
- About 72 cents per video hour by Nvidia estimate: Calculated from an assumed $10 per GPU-hour and modelled throughput. On the same episodes, GPT-6 Astra is quoted at roughly $88.49 per video hour by API. The figure is an estimate excluding storage and overhead.
- Already running in GR00T training: Labelling up to 200,000 hours of video per day across 512 GB300 GPUs. The three models were trained on fewer than 10,000 hours of human and robot video combined.
Use Cases
- Action segmentation and captioning for robot imitation datasets
- Process reward labelling for reinforcement learning
- Execution monitoring and outcome scoring after deployment
- Retrieving specific action clips from long recordings by description
Pros
- All three label types come from one suite, so formats line up without glue code
- The rewarder works at half-second granularity and separates partial completion from outright failure
- Clear division of labour: the 9B model handles segmentation, the 27B models handle captioning and scoring
- Published cost and throughput figures make it easier to size your own deployment
- All three models and the inference code are public on Hugging Face
Pricing
Weights and inference code are public on Hugging Face under the NVIDIA OneWay Noncommercial License, limited to research and evaluation, with derivatives inheriting the restriction. Self-hosting costs depend on your own hardware and throughput. The $0.72 per video hour figure is an estimate built on an assumed $10 GPU-hour rate and modelled throughput, excluding storage and overhead, and it covers captioning only.
Summary
Flora tackles the dullest and most expensive step in robot training: turning video into structured training signal. Three fine-tuned Qwen3.5 models split the job into segmenting, describing and scoring, and NVIDIA says it already runs inside the GR00T training loop.
Two caveats come first. The cost number is a modelled estimate rather than a production bill, and it covers captioning rather than the rewarder. The licence is noncommercial, so research and evaluation are fine while a commercial pipeline needs legal review first. Miss either and you will overestimate how far it reaches.
Version History
- Flora robot video labeling models released (2026-10-25): NVIDIA GEAR lab releases Flora, three Qwen3.5-based models covering action segmentation, timestamped captioning and reward scoring. Already used to train GR00T models, labelling up to 200,000 hours of video per day on 512 GB300 GPUs. Weights and inference code are public on Hugging Face under the OneWay Noncommercial License. Nvidia estimates $0.72 per video hour for captioning.