A Look At The License Behind Nvidia Flora’s Lower-Cost Robot Video Labeling

🔍 Read the full analysis: A Look At The License Behind Nvidia Flora’s Lower-Cost Robot Video Labeling on ThorstenMeyerAI.com

Before you orderOffer from Amazon

Get hardware and tech essentials delivered free with Prime

• Fast, free delivery on millions of items
• Prime Video, Amazon Music and more included
• Member-only deals all year

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Nvidia’s GEAR robotics lab has released Flora, three models designed to segment, describe and score video used to train robots, while licensing them for research or evaluation only. Nvidia estimates Flora can label a video hour for $0.72, compared with $88.49 for GPT-6 Astra, but the Flora figure is a modeled GPU cost rather than a measured production bill, and the license does not authorize ordinary commercial use.

The Flora release comprises three models built on Alibaba’s Qwen3.5: Flora-TempSeg, a 9-billion-parameter model that identifies the time boundaries of actions; a 27-billion-parameter captioner that writes an instruction for each segment; and a 27-billion-parameter reward model that scores robot performance. Nvidia’s model cards say Flora was used as a data engine for its GR00T-2 robot models and was trained on fewer than 10,000 hours of human and robot video.

The cost comparison comes from a test of 50 episodes totaling about 75 minutes in Nvidia’s DenseCaptionBench. Nvidia reports that GPT-6 Astra cost $88.49 per video hour at public API rates, while Flora’s estimated cost was $0.72 per hour, based on eight GB300 GPUs priced at an assumed $10 per GPU-hour. Dividing those figures yields about 122.5, rounded in Nvidia’s presentation to approximately 125 times lower cost. The paper also gives Flora cost scenarios from $0.51 to $1.26 per hour and says storage and other overhead are excluded.

Quality results were mixed rather than a clean win for Flora. On the same episodes, the paper reports a score of 57.1% for Flora and 53.7% for Astra, with lower results for the other systems listed. Flora performed better on some human-video, action-boundary and partial-success evaluations, while Astra led on robot-video caption accuracy and rubric scoring. Most tests were run by Nvidia, and the cost calculation applies to captioning; the reward model is not included in the 125-times comparison.

The License Sets a Commercial Boundary

For robotics companies, the distinction between a model that can be downloaded and one that can be used in a product is central. The NVIDIA OneWay Noncommercial License permits research or evaluation, according to the source material, and requires derivatives to carry the same restriction. That makes Flora available for studying and testing, but not automatically a no-cost labeling service for commercial robot development.

The release still matters as a potential research benchmark and technical reference. Teams can assess whether dedicated video models offer a practical alternative to paying for frontier-model API labeling, subject to the license and hardware demands. Nvidia’s low-cost estimate, if reproduced under comparable conditions, points to one possible way to reduce the expense of preparing robot-training data. It does not settle the wider costs of collecting footage, operating robots safely or deploying suitable compute.

The headline cost gap needs careful interpretation. Astra’s number is reported API usage, while Flora’s is a projection using an assumed GPU-hour price and modeled throughput. Nvidia notes that its production GPU-hour denominator has not been verified. The comparison therefore indicates a possible cost advantage under stated assumptions, not a verified operating cost for a commercial service.

How Flora Labels Robot Footage

Robot training footage can show a gripper placing a mug, two robotic hands passing a wire or a person folding a towel. Models trained on these recordings need labels that identify what action happened and when. The Flora paper describes timestamped captions for filtering and organizing data, step-level rewards that mark normal, suboptimal or failed behavior, and scores for an entire attempt.

Flora divides those tasks across its three models. The captioner can process one camera view or combine a main camera with wrist cameras, while the rewarder can grade short intervals or apply a user-provided rubric. The model cards say the inference code is provided in the Hugging Face repositories; the reported test setup used an Nvidia H100 running Linux. The source estimates that each 27-billion-parameter model needs roughly 55 GB for weights in 16-bit precision, putting use beyond typical consumer-computer capacity.

Nvidia’s benchmark used 20 episodes showing human hands, 15 with two-finger grippers and 15 with dexterous robot hands. The paper reports Flora at 50.3 on the public WGO-Bench segmentation evaluation, compared with Astra at 46.0, and a higher score on the human-video subset of Nvidia’s benchmark. Those results describe particular tests; they do not establish that Flora is better across all robot footage or labeling tasks.

Cost and Performance Still Need Validation

The source material does not establish Flora’s cost at sustained production scale. Its estimate uses an assumed $10 per GPU-hour and modeled throughput, while the Astra figure comes from reported API usage. Nvidia also says the production GPU-hour denominator has not been verified; the estimate excludes storage and overhead. The test covers roughly 75 minutes of video selected for Nvidia’s benchmark, so its results may not represent other datasets or workloads.

Performance also varies by task. Flora’s aggregate result on the reported test was higher than Astra’s, but Astra scored better on robot-video caption accuracy and rubric scoring. The materials do not establish how the models compare across a broader, independently run set of robot deployments. The exact terms governing any permissions beyond research and evaluation, including whether commercial arrangements are available, are not specified in the source material.

Independent Tests and Licensing Terms

The next useful evidence would be independent replication of Flora’s cost and quality results on larger, varied video collections, with GPU prices, throughput and overhead reported clearly. That would help robotics teams compare like with like rather than weigh a modeled GPU budget against an API bill from a limited test.

Organizations considering Flora for product development will also need to review the full license and determine whether separate commercial authorization is available. Until those terms and production costs are clearer, the release is best understood as a research and evaluation resource, not a confirmed commercial labeling pipeline.

Key Questions

What is Nvidia Flora?

Flora is a set of three video models from Nvidia’s GEAR robotics lab. They identify action segments, write captions for them and score robot performance.

Is Flora available for commercial use?

The source material says all three models use the NVIDIA OneWay Noncommercial License, which allows research or evaluation only. Derivatives inherit that restriction; commercial use is not authorized by those terms as described.

Is Flora proven to cost $0.72 per video hour?

No. Nvidia’s $0.72 per-hour figure is modeled using assumed GPU pricing and throughput. The paper says the production GPU-hour denominator has not been verified, and the estimate excludes storage and overhead.

Did Flora outperform GPT-6 Astra?

Results depended on the task. Nvidia reports a higher overall score for Flora on its 50-episode test, while Astra performed better on robot-video caption accuracy and rubric scoring. The reported benchmark does not establish a general winner for every use case.

What happens next for companies evaluating Flora?

Companies will need to check the license terms before using the models beyond research or evaluation, and assess hardware, operating costs and performance on their own data. Independent testing and clarity about any commercial permissions would help inform those decisions.

Source: ThorstenMeyerAI.com