NVIDIA Opensources Robot Data Engine Flora, with Labeling Cost Just 1/125 of GPT-6 Astra
一天能肝20万小时机器人数据，英伟达最新开源数据引擎Flora，成本是GPT-6 Astra零头
NVIDIA has released and open-sourced the robot data engine Flora, which transforms vast amounts of robot video data into high-quality training data using three models: action description, time segmentation, and reward models. Flora performs exceptionally well across multiple benchmarks, with labeling costs just 1/125 of GPT-6 Astra's, and can process 200,000 hours of video data per day.
NVIDIA's Flora data engine demonstrates significant cost efficiency and performance in processing robot video data, achieving results comparable to GPT-6 Astra while using 125 times less cost. The system's modular design, combining action description, time segmentation, and reward models, enables scalable and flexible data annotation for training robot models.
Source: Chinese robotics — Dataset and collection discovery · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
data annotation hours per day
200,000 hours
cost per hour of source video
0.72
View original evidence
具体来看，Flora的标注质量得分为 57.1% ，估算成本为 每小时源视频0.72美元； GPT-6 Astra的标注质量得分为 53.7%， 估算成本为每小时源视频 88.49美元。
Open source S7cost per hour of source video
88.49
View original evidence
具体来看，Flora的标注质量得分为 57.1% ，估算成本为 每小时源视频0.72美元； GPT-6 Astra的标注质量得分为 53.7%， 估算成本为每小时源视频 88.49美元。
Open source S7Reported accuracy
64.7%
View original evidence
同时，使用公开的纯文本动作描述评分标准，联合匹配则要求预测的动作时间区间，和动作描述均与标注结果相符 Flora的平均分段F1为 50.3%， 平均联合F1为 34.5%， 在人工标注参考区间上的动作描述准确率为 64.7%。
Open source S22
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
What remains unknown
Not established in the collected evidence: Environment, Control, Data origin.
Reported performance applies to the described task. It does not establish general autonomy or deployment readiness.
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
目前，该数据引擎依托 512块GB300 GPU， 每天可标注 20万小时的数据。
Open source S8
具体来看，Flora的标注质量得分为 57.1% ，估算成本为 每小时源视频0.72美元； GPT-6 Astra的标注质量得分为 53.7%， 估算成本为每小时源视频 88.49美元。
Open source S7
同时，使用公开的纯文本动作描述评分标准，联合匹配则要求预测的动作时间区间，和动作描述均与标注结果相符 Flora的平均分段F1为 50.3%， 平均联合F1为 34.5%， 在人工标注参考区间上的动作描述准确率为 64.7%。
Open source S22
Source：Chinese robotics — Dataset and collection discovery · sohu.com