HABIT (Human-Aware Behavior and Interaction Training dataset)
收藏资源简介:
HABIT数据集是由KAIST研究团队创建的大规模机器人演示数据集,专门针对人类在场环境设计,旨在促进具身智能中的人机交互研究。该数据集包含10,563个演示片段,总计164小时的双臂操作数据,覆盖60个任务,数据来源于多视角同步摄像头采集的视觉观察与动作序列。数据收集过程通过精心设计的任务工作流和反应式交互协议,系统化地诱发了让步避让、时空同步和手势理解等人类感知行为。该数据集主要应用于训练具有社会兼容性的机器人策略,解决传统人类缺席数据无法编码人机协调行为的关键局限,为部署到真实共享工作空间的机器人提供安全可靠的行为基础。
The HABIT dataset is a large-scale robotic demonstration dataset created by the KAIST research team, specifically designed for human-present environments, aiming to advance human-robot interaction (HRI) research in embodied intelligence. It contains 10,563 demonstration segments, totaling 164 hours of dual-arm manipulation data covering 60 distinct tasks. The data is collected via synchronized multi-view camera recordings of visual observations and action sequences. During the data collection process, through meticulously designed task workflows and reactive interaction protocols, human perceptual behaviors such as concession and avoidance, spatiotemporal synchronization, and gesture understanding are systematically elicited. This dataset is primarily applied to train socially competent robotic policies, addressing the critical limitation that traditional human-absent data cannot encode human-robot coordination behaviors, providing a safe and reliable behavioral foundation for robots deployed in real shared workspaces.

- 1HABIT: Human-Aware Behavior and Interaction Training Dataset for Robot ManipulationKAIST · 2026年



