NVIDIA's GEAR robotics lab, co-led by Jim Fan, has open-sourced Flora, a model suite that turns raw human and robot video into the labels robot foundation models need. The models write timestamped action captions for filtering and balancing pretraining data, dense progress rewards for quality control and reinforcement learning post-training, and rubric scores for real-world evaluation. NVIDIA already uses the pipeline to train its GR00T models, and the weights let other labs run that annotation step on their own footage. The lab reports that Flora is competitive with GPT-6 Astra on internal benchmarks at roughly 125 times lower cost, and that production labeling covers 200,000 hours of video a day across 512 GB300 GPUs. NVIDIA also reports state-of-the-art scores on the captioning benchmark WGO-Bench and the reward benchmarks RoboReward and Robometer. The release is for researchers scaling custom robotics datasets, including human-to-robot transfer.