OpenGVL
收藏资源简介:
OpenGVL是一个全面的基准测试,用于评估在各种具有挑战性的操作任务中估计任务进度。它涉及机器人和人类的双重体现,旨在通过视觉观察预测任务进度。OpenGVL评估了公开可用的开源基础模型的能力,并展示了开源模型家族在时间进度预测任务上的性能显著低于闭源模型。OpenGVL还可以作为自动化数据管理和过滤的实用工具,以有效地评估大规模机器人数据集的质量。
OpenGVL is a comprehensive benchmark for evaluating task progress estimation across various challenging manipulation tasks. It features dual embodiments of both robots and humans, aiming to predict task progress via visual observations. OpenGVL evaluates the capabilities of publicly available open-source foundational models, and demonstrates that the performance of open-source model families on temporal progress prediction tasks is significantly inferior to that of closed-source models. Additionally, OpenGVL can serve as a practical tool for automated data management and filtering to efficiently assess the quality of large-scale robotic datasets.




