FLORA: An Open Model Suite for Scaling Robotics Data Engines

At a glance

Problem: most labeling systems cover only one annotation type, and general-purpose frontier models are expensive at robot-dataset scale. Source
Idea: stitch overlapping segmentation windows into event intervals, caption each interval with neighboring context, and use one reward model for dense, outcome or rubric labels. Source
Evidence: on RoboRewardBench the reward model's outcome error (MAE, multiplied by 100) is 40.0, versus 55.4 for GPT-6 Astra and 66.5 for the specialized RoboReward model. Source

Evaluation scope Reward results combine two public and two internal benchmarks; numbers for specialized reward models are taken from their own publications under their original protocols. Source

Flora's capabilities: videos are segmented and each segment captioned, and the same suite produces outcome, process and rubric-based rewards. Original figure

How it works

The segmentation model predicts window-relative start and end times with continuation flags; fragments are converted to episode time and joined using temporal overlap and those flags. Source
The captioning model writes one imperative instruction per interval from the target frames plus labeled before and after context, and failed attempts keep their intended goal without implying success. Source
Flora-Rewarder labels each 0.5-second step as failure, suboptimal or normal, and the labels are retrospective: a later recovery does not improve an earlier step. Source
The segmenter is fine-tuned from Qwen3.5-9B with LoRA and a frozen visual encoder, while the captioner and reward model start from Qwen3.5-27B. Source
The supervised segmenter is further trained with GRPO using a verifiable reward built from temporal F1, precision and endpoint accuracy. Source

Evidence & results

On DenseRewardBench (40 held-out bimanual episodes) Flora-Rewarder recalls 53.3% of failure steps while no baseline exceeds 11.7%, but it also labels about 11% of normal steps as failures. Evidence
On the authors' DenseCaptionBench, Flora leads all three metrics on Human100 (37.4% joint F1) but ranks second to GPT-6 Astra on Robot200 (51.0% versus 57.9% joint F1). Evidence
On WGO-Bench pipelines, Flora improves mean segmentation F1 by 16.9 points and joint F1 by 7.3 points over Refiner, with similar caption accuracy (64.7% vs. 64.3%). Evidence

Caption and reward labeling results across six tasks, with all values scaled by 100 as in the paper's figure. Original figure

Limitations & boundaries

On the in-house rubric benchmark GPT-6 Astra is more accurate, with a pooled L1 of 3.6 versus Flora-Rewarder's 6.5; the authors call these system-level comparisons. Source
Caption scores are judged by GPT-6 Astra, which is also an evaluated system, so the authors flag generator–judge self-preference as a limitation. Source
Dense labels use later frames as evidence, so the authors intend these retrospective annotations for offline labeling. Source

Takeaway

Our reading: Flora's strongest case is labeling cost, since on a 50-episode subset it scores 57.1% quality versus 53.7% for GPT-6 Astra at an estimated $0.72 rather than $88.49 per hour of video, an estimate based on the authors' own GPU deployment assumptions. Source

Resources

Full text ↗PDF ↗

Figures: the paper's authors. Editorial source and image review: 2026-10-26. Findings are author-reported, not independently reproduced by Robotics Research Hub.

Original 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.

Citation

@article{Leng2026_261031779,
  title     = {FLORA: An Open Model Suite for Scaling Robotics Data Engines},
  author    = {Sicong Leng and Yunze Man and Zhiqi Li and Jing Wang and Yuhe Zhang and Ryan Punamiya and Siddharth Patki and Kaiyuan Zheng and Zi-ang Cao and Zhiding Yu and Sifei Liu and Danfei Xu and Ben Levin and Jan Kautz and Yuke Zhu and Johan Bjorck and Linxi "Jim" Fan},
  year      = {2026},
  eprint    = {2610.31779},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url       = {https://arxiv.org/abs/2610.31779}
}