Recently, NVIDIA's GEAR research lab formally open-sourced its open data engine Flora. Built on three fine-tuned Qwen3.5 models , the system aims to provide efficient text annotation and scoring support for robot video training, achieving a key positioning across the robotics value chain.
Flora can comprehensively cover the three core label types needed for robot training: action boundary detection, step-by-step execution scoring, and task success judgment. Its internal integration of an action segmentation model and the reward model Flora-Rewardersignificantly improves the ability to capture mid-course errors and detailed actions while lowering costs.
According to NVIDIA's estimates,Flora 's annotation cost is approximately 0.72 US dollars per hour, compared to about 88.50 US dollars for GPT-6Astra approximately US$88.50, cutting costs substantially by up to 125 times. In terms of actual effectiveness, it performs excellently in action boundary detection and process error identification, and leads in human videos and some benchmarks, though it still lags behind in pure description accuracy and performance on some robot videos GPT-6Astra。
in pure captioning accuracy and performance on some robot videos. As an important move by NVIDIA to strengthen its embodied AI ecosystem, the model is already live on Hugging Face , and internally processes up to 200,000 hours of video annotation per day. This move not only fills the gap in open-source robot data pipelines but also further consolidates NVIDIA's industry strategy of full-chain coverage from hardware and simulation to AI models.