Nvidia open-sources Flora for robotics video labelling

Nvidia has open-sourced Flora, an AI data engine designed to label the videos used to train robots. The company says the model suite can produce timestamped descriptions, process rewards and outcome rewards in one pipeline, while costing roughly 125 times less than GPT-6 Astra in its internal comparison. ¹

That matters because robotics systems need far more than a video archive. Training data must explain what a robot is doing, when an action starts and ends, whether each step is going well and whether the overall task succeeds. Flora is intended to automate those judgements at a scale that would be difficult to sustain with paid, general-purpose AI services.

What Flora does

Nvidia's GEAR research team says Flora covers three jobs. Its captioning models split footage into timed actions and describe what happens in each segment. Its reward models score progress through an episode and judge whether a task has been completed successfully. The combined output can be used to filter training data, monitor deployments and support reinforcement learning. ¹

The open suite is built from three fine-tuned Qwen3.5 models. A 9B model identifies action segments, while a 27B model describes those actions using the surrounding context. A second 27B model, Flora-Rewarder, can score success, classify individual steps as normal or failed and apply a rubric to a trajectory. The models and supporting material are available through Nvidia's Flora release and Hugging Face collection. ¹²

The performance and cost claims

Nvidia reports that Flora is competitive with GPT-6 Astra on internal benchmarks and achieves leading results on several of its published comparisons. On WGO-Bench, it is particularly strong at locating the boundaries between actions, although Astra can produce more precise descriptions in some cases. The research also reports a lower error rate on partially completed robot tasks, where a system must recognise that a movement has made progress without finishing the job. ¹

The headline cost comparison is an estimate rather than a retail price or a production invoice. Nvidia calculates that one hour of video labelling costs about 72 cents with Flora, compared with roughly $88.50 for GPT-6 Astra, based on an assumed GPU rental cost of $10 per hour. That makes the 125-times figure useful as an indication of the potential gap, but not as a guarantee of what every organisation will pay.

The paper also records trade-offs. Flora can split a long movement into too many short actions, and it occasionally misidentifies what a robot is doing. Nvidia's own results show that different systems lead on different measures: one model may find action boundaries better while another writes a more accurate description. Those caveats are important for researchers deciding whether an open model is ready for a particular dataset.

Why Nvidia is open-sourcing it

Nvidia says Flora is already used during the training of its GR00T robotics models, processing up to 200,000 hours of footage per day across 512 GB300 GPUs. Releasing the models gives researchers a way to build larger robotics datasets without sending every clip to a commercial API, and it lets them test the same types of captioning and reward signals on their own hardware. ¹

It also fits Nvidia's wider strategy. The company is supplying chips, robot foundation models and the software used to turn raw footage into training material. An open data engine can encourage more experimentation around GR00T while keeping Nvidia hardware central to the most demanding workloads. That is useful for developers, but it is also a clear ecosystem play rather than a purely neutral research giveaway.

Our opinion

Flora is most interesting as infrastructure, not as another chatbot-style model launch. Robotics teams do not just need a model that can describe a single clip; they need repeatable labels across huge, messy datasets. A system that can identify actions and failures at scale could remove one of the less glamorous bottlenecks in robot development.

The release should still be read with the same caution as any vendor benchmark. Nvidia supplies the cost assumptions and several of the comparisons, and the research acknowledges that Flora is not best at every kind of description. Its real test will be whether independent teams can reproduce the results on unfamiliar robots and whether the open models remain useful when the videos are less tidy than a benchmark set.

For now, the practical story is straightforward: Nvidia has made a specialised robotics labelling stack available to researchers, with a strong claim about cost and a more nuanced set of claims about accuracy. That could make open experimentation cheaper, even if the company remains closely tied to the infrastructure needed to run it.