Nvidia has open-sourced Flora, a system that turns robot video into three kinds of training labels: timestamped action descriptions, step-by-step progress ratings and judgments about whether an entire task succeeded. The company says the pipeline costs about 72 cents per video hour, compared with $88 using GPT-6 Astra.
Flora uses three fine-tuned Qwen 3.5 models. A 9-billion-parameter model splits footage into action segments. One 27-billion-parameter model describes those segments using surrounding context, while another scores overall success and classifies each half-second step as normal, suboptimal or failed.
Nvidia’s GEAR research lab says it already uses Flora while training GR00T robotics models, processing as much as 200,000 hours of footage per day on 512 GB300 GPUs. The models are available through Hugging Face. In the reported evaluation, Flora slightly exceeded the general-purpose model’s aggregate quality while costing about 125 times less.
The cost comparison needs context. Specialized local infrastructure has utilization and engineering costs beyond a per-inference estimate, and the article notes that Flora can be less accurate on some descriptive labels even while detecting action failures well. Results from Nvidia’s own tests also need independent reproduction. The release is still useful because robotics teams can inspect and adapt the models rather than sending every video to an expensive closed service.