Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra

Nvidia has released Flora as an open-source data engine designed to label robot video data at dramatically lower cost than frontier LLMs. The system is built on three fine-tuned Qwen3.5 models and reportedly labels one hour of robot video for 72 cents, compared to $88 when using GPT-6 Astra, a roughly 125x cost reduction.

According to the report, Flora also scores slightly higher on labeling quality than the GPT-6 Astra baseline and catches errors in robot actions more reliably than competing approaches. This combination of lower cost and higher accuracy makes it a notable entry in the robot data labeling space.

Robot video labeling is a critical bottleneck in training embodied AI systems. Historically, teams have relied on expensive frontier LLMs to annotate robot action sequences, which has limited the scale of training data many organizations can afford to process.

By open-sourcing Flora, Nvidia lowers the barrier for robotics researchers and developers to build large-scale labeled datasets. The decision to base the system on Qwen3.5 fine-tunes rather than proprietary Nvidia models also signals a pragmatic approach to leveraging open-weight foundations for specialized data pipelines.

The release could accelerate progress in embodied AI by making high-quality robot training data more accessible, though real-world performance at scale will need independent validation beyond the reported benchmarks.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.