NVIDIA open-sources Flora: robot video labeling suite claimed ~125× cheaper than GPT-6 Astra
NVIDIA GEAR released Flora—three Qwen3.5 fine-tunes (TempSeg ~9B LoRA, Captioner 27B, Rewarder 27B) for timestamped action captions plus process/outcome rewards (optional rubrics). NVIDIA says it feeds GR00T training at up to ~200k video-hours/day and estimates ~$0.72 vs ~$88.50 per video-hour versus GPT-6 Astra. Weights are on Hugging Face under the NVIDIA OneWay Noncommercial License (research/dev only).
NVIDIA’s GEAR lab published Flora, an open data engine meant to cover three label types in the robotics foundation-model lifecycle: timestamped action captions for filtering/balancing pretraining data, process rewards for RL post-training, and outcome rewards for real-world monitoring—plus optional test-time rubric scores. NVIDIA says Flora is already used while training GR00T models and has labeled on the order of 200,000 video hours per day across 512 GB300 GPUs.
The suite is three vision-language models fine-tuned from Qwen3.5: Flora-TempSeg / Flora-segmenter-9B (~9B params, LoRA on attention and MLP projections with rank 32 and alpha 64; base and vision encoder frozen) cuts action intervals; Flora-Captioner 27B writes one action instruction per interval with neighboring context; Flora-Rewarder 27B (vendor: full fine-tune including the vision encoder) emits outcome scores, dense process labels, or rubric scores by prompt. Hugging Face model cards list a 10/15/2026 release and the NVIDIA OneWay Noncommercial License (22 March 2022)—research and development only, not commercial use or sole safety monitoring of physical robots.
On cost, the project page and paper call Flora roughly ~125× cheaper than GPT-6 Astra. THE DECODER restates NVIDIA’s ~$0.72 vs ~$88.50 per source-video-hour figures from a ~50-video trial and notes they are projections assuming about $10 per GPU-hour rent, not metered production bills. On quality, NVIDIA claims leads or near-parity with Astra on several internal/public reward settings while trailing on some captioning axes such as parts of WGO-Bench; treat the scores and the “125×” line as NVIDIA claims. The paper PDF is dated 2026-10-20 in the header; Decoder’s public write-up is 2026-10-25.
Sources
- Flora: An Open Model Suite for Scaling Robotics Data Engines — NVIDIA Research / GEAR
- FLORA: An Open Model Suite for Scaling Robotics Data Engines — NVIDIA Research
- nvidia/Flora-TempSeg · Hugging Face — Hugging Face / NVIDIA
- nvidia/Flora-Captioner · Hugging Face — Hugging Face / NVIDIA
- nvidia/Flora-Rewarder · Hugging Face — Hugging Face / NVIDIA
- Flora - a nvidia Collection — Hugging Face / NVIDIA
- Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra — THE DECODER