NVIDIA GEAR releases Flora, open models that caption and score robot and human videos to train GR00T; weights are non-commercial

Why it matters

【Key points】

NVIDIA's GEAR lab released Flora on Oct 20: open-weight models that label robot training data. They watch human and robot videos, write timestamped action descriptions and score how well each attempt went. NVIDIA says it uses Flora to train its GR00T robot models.

1. Three models: Flora-Captioner (~27B, from Qwen3.5-27B) writes timestamped captions; Flora-Rewarder (~27B, from Qwen3.5-27B) gives episode-level success rewards and dense process rewards for RL post-training and deployment monitoring; Flora-TempSeg is a Qwen3.5-9B LoRA adapter that finds where each action starts and ends.

2. Scale (NVIDIA's figures): it labels 200,000 hours of video a day on 512 GB300 GPUs, at roughly 1/125 the cost of GPT-6 Astra.

3. Results (NVIDIA's tests): NVIDIA claims state of the art on WGO-Bench, RoboReward and Robometer and parity with GPT-6 Astra on internal tests. Its table shows them trading wins: Flora leads on human-video captions and failure detection, Astra on robot-video captions and several WGO-Bench metrics. On most metrics both are well ahead of GPT-5.6 Sol, Claude Opus 5 and Kimi K3.

4. License: NVIDIA OneWay Noncommercial, research and development only.

【Verifiable】

Project page: https://research.nvidia.com/labs/gear/flora/

Collection: https://huggingface.co/collections/nvidia/flora

Captioner: https://huggingface.co/nvidia/Flora-Captioner

Rewarder: https://huggingface.co/nvidia/Flora-Rewarder

TempSeg: https://huggingface.co/nvidia/Flora-TempSeg

Jim Fan post: https://x.com/DrJimFan/status/2112597587987689912

Author post: https://x.com/yunzeman/status/2112590210840281500

【Our read】

1. Turning video into training signal is a robot-learning bottleneck; Flora lets labs label and score their own egocentric and robot data.

2. "Near Astra at 125x lower cost" is NVIDIA's own comparison and still needs outside replication.

3. The non-commercial license limits it to research; startups cannot ship it in products.