Flora: An Open Model Suite for Scaling Robotics Data Engines

TL;DR

NVIDIA GEAR open-sources Flora, an industrial-strength robotics data engine that unifies three distinct label types — timestamped captions, dense per-frame reward annotations, and rubric scores — behind one open-weight model suite. Flora already powers the in-house pipeline training NVIDIA GR00T, labeling 200,000 hours of robot and human-egocentric video per day across 512 GB300s. On GEAR’s internal benchmarks it is competitive with GPT-6 Astra while beating other frontier models (GPT-5.6 Sol, Claude Opus 5, Kimi K3) and is roughly 125× cheaper than Astra per unit of labeled footage. The pitch is that Flora closes the frontier-VLM annotation gap for open robotics foundation models the same way DeepSeek-OCR closed the document-understanding gap earlier in 2026 — one open model suite covering captions, process rewards, and outcome rewards simultaneously.

Key claims

Flora achieves SOTA scores on timestamped-caption, dense-reward, and rubric-scoring benchmarks simultaneously — WGO-Bench, DenseCaptionBench (Human100 + Robot200 subsets), RoboRewardBench, RobometerBench, DenseRewardBench, and RubricBench — and the per-benchmark numbers beat Claude Opus 5 and Kimi K3 across all three categories while trading places with GPT-6 Astra and GPT-5.6 Sol depending on the metric [§Benchmark table].
Flora matches or beats GPT-6 Astra on timestamped captions for human-egocentric video — DenseCaptionBench Human100 subset: segmentation F1@0.5 54.6 vs 44.2, caption point quality 66.5 vs 60.3, joint point F1@0.5 37.4 vs 29.7 — while trading places with Astra and GPT-5.6 Sol on robot-camera video (Robot200 subset Astra wins caption point quality 73.7 vs 80.2) [§Benchmark table].
Flora wins on WGO-Bench segmentation F1@0.75 (50.3 vs Astra 46.0, GPT-5.6 Sol 33.7), the Macrodata Labs benchmark on robot/egocentric subtask boundary discovery, while GPT-5.6 Sol wins the gold-window-accuracy sub-metric (77.4 vs Flora 64.7) [§Benchmark table].
On dense reward annotations, Flora beats Astra on RoboReward MAE (40.0 vs 55.4, lower is better) and all four frontier models on DenseRewardBench macro F1 (46.2 vs 40.7 Astra, 36.3 Sol, 33.9 Opus 5, 31.2 Kimi K3), with the sharpest gap on failure recall (53.3 vs 0.0 Astra, 1.0 Sol, 0.3 Opus 5, 0.3 Kimi K3) [§Benchmark table].
On Robometer OOD Kendall (out-of-distribution rank correlation), Astra and Sol beat Flora (82.3 Sol, 72.0 Flora, 64.2 Opus 5) — Astra is scored lower (not reported in the OOD row shown) but Sol holds the lead; in-distribution and OOD VOC are near-ceiling for all four frontier systems [§Benchmark table].
On RubricBench pooled L1 error (lower is better), GPT-5.6 Sol wins (3.6) with Flora second (6.5), then Astra (8.4), Opus 5 (17.1), Kimi K3 (15.6) — the only sub-category where GPT-5.6 Sol clearly out-ranks Flora [§Benchmark table].
Flora is roughly 125× cheaper than GPT-6 Astra per unit of labeled video [§TL;DR / production run figure], the sharpest filed cost-per-hour datapoint for frontier-grade robotics annotation at open-weight scale.
Deployment proof: Flora labels 200,000 hours of robot and human egocentric video per day on a 512 GB300 production cluster, directly feeding GR00T training [§Production run figure].
The model suite covers the three label types a robotics foundation-model lifecycle depends on: (a) timestamped captions for filtering and balancing pretraining data; (b) dense process-reward labels for RL post-training; (c) outcome-reward and rubric labels for monitoring real-world deployments [§Opening statement].

Method

Flora is presented as a suite of open-weight models unified behind one annotation pipeline rather than a single model. The project page states that “human and robot” are handled by the same annotation pipeline — the three subsets of DenseCaptionBench (Human100 for egocentric, Robot200 for robot-camera) are evaluated with the same Flora backbone. The release includes model weights on Hugging Face (collection nvidia/flora) and the project page positions the suite as an open substrate for both internal GEAR training (GR00T) and external research use.

Precise architecture details, pretraining corpus, and inference-time cost structure are not disclosed in the project page or Jim Fan’s launch tweet at filing time. The announcement surfaces three capability promises — timestamped captions, dense progress rewards (for quality control and RL post-training), and rubric scoring (for real-world benchmarking) — framed as the three outputs any robotics data engine needs.

A citation is already in place: @article{leng2026flora, title = {Flora: An Open Model Suite for Scaling Robotics Data Engines}, author = {Sicong Leng and Yunze Man ... Linxi "Jim" Fan}, year = {2026}}, so a technical paper is implied but not published alongside the project page.

Results

Internal benchmark comparison table (↑ higher better, ↓ lower better; abridged from the Flora project page):

DenseCaptionBench Human100 segmentation F1@0.5 ↑: Flora 54.6 / Astra 44.2 / Sol 32.6 / Opus 5 22.7 / Kimi K3 21.1.
DenseCaptionBench Robot200 segmentation F1@0.5 ↑: Flora 66.9 / Astra 71.9 / Sol 45.8 / Opus 5 45.9 / Kimi K3 41.2.
WGO-Bench segmentation F1@0.75 ↑: Flora 50.3 / Astra 46.0 / Sol 33.7 / Opus 5 25.5 / Kimi K3 26.1.
WGO-Bench gold-window accuracy ↑: Flora 64.7 / Astra 77.4 / Sol 56.6 / Opus 5 49.3 / Kimi K3 50.6.
RoboReward MAE ↓: Flora 40.0 / Astra 55.4 / Sol 78.0 / Opus 5 88.2 / Kimi K3 68.2.
Robometer OOD Kendall ↑: Flora 72.0 / Sol 82.3 / Opus 5 64.2 / Kimi K3 63.5 (Astra value not shown in the extracted row).
Robometer ID VOC ↑: Flora 98.9 / Astra 98.9 / Sol 97.1 / Opus 5 91.8 / Kimi K3 97.2.
DenseRewardBench macro F1 ↑: Flora 46.2 / Astra 40.7 / Sol 36.3 / Opus 5 33.9 / Kimi K3 31.2.
DenseRewardBench failure recall ↑: Flora 53.3 / Astra 0.0 / Sol 1.0 / Opus 5 0.3 / Kimi K3 0.3.
RubricBench pooled L1 error ↓: Flora 6.5 / Sol 3.6 / Astra 8.4 / Opus 5 17.1 / Kimi K3 15.6.
Production throughput: 200,000 hours of video per day on 512 GB300 GPUs [§Production run figure].
Cost: ~125× cheaper than GPT-6 Astra per unit of labeled footage [§TL;DR / production run figure].

Why it’s interesting

Flora is the first open-weight model suite that directly targets the three label types the wiki has been triangulating from multiple directions — captions, dense process rewards, and rubric scores — in one release. Argus: An Open-Source Annotator for Robotics Data is the closest filed sibling on the pipeline layer: Pantheon’s Argus audit of five frontier VLMs across 193 robotics episodes shows Astra at 44 events/min with the three mid-tier peers ≤23/min and GPT-6.1 Sol reaching 43/min at $5 per hour of footage vs Astra's$ 27. Flora now claims to beat Astra on dense-reward failure recall by two orders of magnitude (53.3 vs 0.0) at ~125× lower cost — the same open-vs-frontier gap Argus measured, but closed via open weights rather than frontier-VLM selection.

Flora is also the direct NVIDIA answer to the WGO-Bench benchmark introduced by Segmenting Robot Video into Actionable Subtasks (WGO-Bench) — Macrodata’s own best end-to-end F1 using Gemini 3.5 Flash with contact-sheet prompting was 0.168 on segmentation at $2.64 per hour of video. Flora reports WGO-Bench segmentation F1@0.75 of 50.3 (a stricter IoU threshold than Macrodata’s 0.75+ grading, so the numbers aren’t directly comparable at the metric level) but the implicit claim is that an open model specifically trained for this task beats Gemini-with-prompting both on quality and (via “125× cheaper than Astra”) on cost. The dense-reward side has the same shape: DenseReward: Dense Reward Learning via Failure Synthesis for Robotic Manipulation trained a VLM-based dense-reward model on simulator-synthesized failure trajectories; Flora now ships the same capability at frontier-matching accuracy on real-robot RoboRewardBench + DenseRewardBench and at production-scale throughput.

Finally, this is Jim Fan’s second major open-model release on this axis in two weeks after ARC: A Reasoning Recipe for Robot Foundation Models (ARC) — both pitched as “open substrate for scaling physical AI” and both backed by NVIDIA’s GB300 infrastructure. The pair sketches a strategy: ship open frontier-matching tools for both the reasoning layer (ARC) and the data engine layer (Flora), and let third parties train their own robot foundation models against the two.