Nvidia has open-sourced Flora to make robot video annotation 125 times cheaper than GPT-6 Astra.

Meta Description: Nvidia Flora uses three Qwen3.5 models to segment, describe, and score robot videos while Nvidia reports it can annotate 200,000 hours daily for GR00T training.

Tags: Nvidia Flora, GR00T, robotics video annotation, Flora-Rewarder, Qwen3.5

Featured Image: Abstract technology style image with segmented video timelines, action boundary markers, reward score rings, and GPU cluster shapes; no real people, no company logos, no text.

According to the paper, Nvidia Flora is an open system that covers three tasks in robotics video review: dividing clips into time segments, describing actions, and evaluating how an episode was executed and ended. Nvidia says the system is already running during training of its GR00T models, annotating up to 200,000 hours of video per day across 512 GB300 GPUs. The models are available on Hugging Face. That matters because the paper describes it as a tool for generating the labels and reward signals that robotics training pipelines need.

Background: Data Workloads Shape the Robotics Stack

Nvidia's broader effort began before Flora: the company has spent some time trying to offset a shortage of robotics training data by using simulation and automated data processing. At GTC earlier this year, Nvidia demonstrated a pipeline that prepares, multiplies, and scores training material automatically. Nvidia has also invested billions in open AI models, with the stated goal of keeping developers inside its hardware ecosystem and sustaining chip demand. Flora supports that goal by adding detailed video labels and reward signals to the pipeline.

Development: Three Qwen3.5 Models Split, Describe, and Score

The system is built from three fine-tuned Qwen3.5 models. A 9B model divides clips into action segments. A 27B model writes descriptions that consider events before and after each action. A second 27B model, Flora-Rewarder, performs three jobs depending on the prompt: it scores episode success from 1 to 5, labels each half-second step as normal, suboptimal, or failure, and applies self-defined rubrics to assign scores.

The design separates perception from judgment: segmentation and captioning produce timestamped descriptions, while one reward model creates all three rating types.

Latest Progress: Cost Projections and Benchmark Tradeoffs

Nvidia reports that, in a 50-video test, Flora edged out GPT-6 Astra even when OpenAI's model used full reasoning capacity. According to Nvidia's estimate, labeling one hour of video costs about 72 cents with Flora, compared with roughly $88.50 for GPT-6 Astra. Claude Opus 5 performed much worse in the same test while costing about $42 per hour.

The reported comparison also places Flora at 57.1 percent annotation quality at $0.72 per video hour, while GPT-6 Astra reaches 53.7 percent at $88.49. By Nvidia's own estimates, that is up to 125x cheaper. The cost figures are projections, not production measurements. Nvidia calculated them based on an assumed rental price of $10 per GPU per hour.

In public WGO-Bench results, Flora performs best at finding action boundaries, but it trails Astra overall and falls behind Gemini 3.8 Flash on description accuracy alone. The researchers say the improvement comes mainly from temporal localization. On Nvidia's DenseCaptionBench, Flora leads every metric for human recordings; for robot videos, it places second behind Astra. The researchers also acknowledge a methodological catch: GPT-6 Astra acts as an evaluator for descriptions while also being one of the models under test.

Flora leads on human recordings, outcome scoring, and step-by-step ratings, while GPT-6 Astra wins on WGO-Bench, robot videos, and rubric scoring. On RoboRewardBench, Flora's error rate is 27.8 percent lower than GPT-6 Astra's. The paper links the gap to partially completed tasks. If a robot only tips a bottle over instead of knocking it off the table, Flora assigns 3 out of 5 points, matching the human reference, while Astra assigns just one.

Flora-Rewarder catches 53.3 percent of failed steps within an episode. No competitor exceeds 11.7 percent, and Astra catches none. The tradeoff is that Flora falsely flags about 11 percent of normal steps as errors. Astra still leads on rubric-based scoring and trajectory ranking.

In a video with many short hand movements, Flora identifies 34 of 37 individual actions, while Astra catches only four. On the flip side, Flora sometimes splits longer actions into too many small steps and occasionally gets descriptions wrong. In one video, a robot hand wipes a plate with a cloth, but Flora describes it as folding a napkin.

Current Stage: An Open Labeling Layer in a Competitive Pipeline

Nvidia positions Flora as an open alternative to commercial services like Scale AI's Physical AI Data Engine. Researchers can use it to scale their own robotics datasets and test approaches like transferring human demonstrations to robots. The strategy aims to cover the entire robotics value chain, from hardware and training data to simulation and the AI models themselves.

The latest competitive reference point is Google's recently announced Gemini 4 Argon, which targets a similar part of the robotics training process. Like Flora, it identifies work steps in training videos and flags failed actions.