flora

The high cost of robot training data has taken another hit from Nvidia.

According to reports, Nvidia has open-sourced its data engine Flora, claiming it can bring the cost of annotating robot videos down from about $88 to around $0.72—roughly 125 times cheaper than the purported GPT-6 Astra approach. The process broadly involves segmenting time periods, describing actions, then evaluating execution and outcomes. It has reportedly been used to train GR00T, and can annotate up to around 200,000 hours of video per day across a large number of GPUs.

How to read this: An open-source entry point lowers the barrier to access, but real-world adoption will still depend on whether annotation quality and gains in downstream models live up to the claims. If the data bottleneck in robotics and embodied AI is eased, training cycles could speed up. Keep an eye on reproducible benchmarks first; don’t get carried away by the “125 times cheaper” headline.

#Nvidia #Flora #机器人 #OpenSource