ember-lab-berkeley/robocasa365-pretrain-composite
收藏资源简介:
该数据集是RoboCasa365的482小时预训练人类数据集的长时部分,包含235个复合(多步骤)任务的人类远程操作数据(每个任务约100个演示)。数据集包括24,687个episodes,27,610,913帧(20 fps,相当于383小时),4,428个任务(自然语言表述,底层RoboCasa任务类别为235个)。摄像机配置为3个256×256 h264视频(robot0_agentview_left / _right / _eye_in_hand),机器人配置为PandaOmron(Panda 7-DoF臂 + Omron移动底座 + 躯干升降)。数据集的特征模式包括观察状态(float64[16])、动作(float64[12])和观察图像(video, 256×256×3)。数据集采用MIT许可证,引用自RoboCasa365论文(ICLR 2026)。
This dataset is the long-horizon half of RoboCasa365s 482-hour pretraining-human dataset, containing human teleoperation data for 235 composite (multi-step) tasks (~100 demos/task). It includes 24,687 episodes, 27,610,913 frames (20 fps → 383 h), and 4,428 tasks (natural-language phrasings; underlying RoboCasa task classes: 235). The camera setup consists of 3 × 256×256 h264 videos (`robot0_agentview_left` / `_right` / `_eye_in_hand`), and the robot is a PandaOmron (Panda 7-DoF arm + Omron mobile base + torso lift). The feature schema includes observation state (float64[16]), action (float64[12]), and observation images (video, 256×256×3). The dataset is licensed under MIT, inherited from upstream RoboCasa365, and cited in the RoboCasa365 paper (ICLR 2026).




