habit-anonymous/HABIT

HABIT是一个大规模机器人演示数据集，专门针对人类在场环境设计，旨在训练机器人策略以学习人类感知行为。该数据集包含48个任务、7,518个片段、4.03百万帧和111.83小时的视频数据，每个片段都有人类共同参与机器人工作空间。数据集定义了三个交互角色：Collaborator（协作角色，涉及直接物理交互）、Coworker（同事角色，共享目标但无直接接触）和Supervisor（监督角色，通过手势或指令指导机器人）。数据收集使用双Franka Research 3机器人手臂和5个RGB摄像头（3个机器人侧、2个人类侧），并包括子任务级注释，记录人类和机器人的子任务边界。数据集采用LeRobot v2.0格式，提供完整版本（约160 GB）和样本版本（约1 GB），适用于微调视觉-语言-动作模型、研究人机交互和协作任务。HABIT is a large-scale robotic demonstration dataset specifically designed for human-present environments, aiming to train robotic policies to learn human-aware behaviors. This dataset contains 48 tasks, 7,518 segments, 4.03 million frames, and 111.83 hours of video data, with each segment featuring human co-participation in the robot's workspace. The dataset defines three interaction roles: Collaborator (a role involving direct physical interaction), Coworker (a role with shared goals but no direct physical contact), and Supervisor (a role that guides the robot via gestures or commands). Data collection was conducted using dual Franka Research 3 robotic arms and five RGB cameras (three mounted on the robot side, two on the human side), and it includes subtask-level annotations that record the subtask boundaries for both humans and robots. The dataset is provided in LeRobot v2.0 format, offering both a full version (approximately 160 GB) and a sample version (approximately 1 GB), which is suitable for fine-tuning vision-language-action models, as well as researching human-robot interaction and collaborative tasks.