Johnbosco20/community_dataset_v3

Lerobot社区数据集v3是一个大规模机器人数据集，专为视觉-语言-动作学习设计。该数据集包含791个子数据集，涵盖46种机器人类型，支持跨体现预训练，适用于通用机器人策略的开发。数据集由全球235名社区贡献者共同构建，并经过清理和组织，确保数据质量。数据集包含单臂、双臂、移动操作和人形机器人等多种机器人类型，总时长251.5小时（10.5天），总帧数25,971,082帧。数据集结构包括观测数据、视频记录和元数据，适用于多种任务和研究，如跨体现VLA训练、多任务操作、迁移学习、模仿学习、通用机器人策略、移动操作和具身AI研究。

Lerobot Community Datasets v3 is a large-scale robotics dataset designed for vision-language-action learning. It features 791 datasets across 46 robot types, enabling cross-embodiment pretraining for generalist robot policies. Compiled from 235 community contributors worldwide, this cleaned and organized dataset includes single-arm, bimanual, mobile manipulation, and humanoid robots, totaling 251.5 hours (10.5 days) of data with 25,971,082 frames. The dataset structure comprises observation data, video recordings, and metadata, supporting various tasks and research such as cross-embodiment VLA training, multi-task manipulation, transfer learning, imitation learning, generalist robot policies, mobile manipulation, and embodied AI research.