ember-lab-berkeley/robocasa365-target-atomic
收藏资源简介:
该数据集名为Target (Human) — atomic-seen,包含人类遥控操作的18个原子任务的演示数据(每个任务500次演示),在10个目标厨房中记录。这些任务也在预训练数据集中有所体现。数据集是RoboCasa365集合的一部分,采用LeRobot v3.0镜像格式,可直接加载。数据集包含9,126个片段,2,231,347帧(20 fps,约31小时),231个任务(自然语言表述,基于18个RoboCasa任务类别)。数据采集使用了3个256×256分辨率的摄像头,机器人设备为PandaOmron(7自由度手臂+Omron移动底座+躯干升降)。
The dataset is named Target (Human) — atomic-seen, containing human teleoperation data for 18 atomic tasks (500 demos/task) recorded in 10 held-out target kitchens. All tasks are also represented in the pretraining datasets. It is part of the RoboCasa365 collection, using a flat LeRobot v3.0 mirror format for drop-in loadability. The dataset includes 9,126 episodes, 2,231,347 frames (20 fps → 31 h), and 231 tasks (natural-language phrasings; underlying RoboCasa task classes: 18). Data was captured using 3 × 256×256 resolution cameras, with a PandaOmron robot (Panda 7-DoF arm + Omron mobile base + torso lift).




