PHASE (PHysically-grounded Abstract Social Events)

PHASE 是一个基于物理的抽象社交事件的数据集，通过包含诸如帮助另一个代理之类的社交概念，类似于广泛的现实生活中的社交互动。 PHASE 由成对的代理在连续空间中移动的 2D 动画组成，这些动画使用物理引擎和分层规划器以程序方式生成。代理具有有限的视野，并且可以在具有多个地标和障碍物的环境中与多个对象交互。使用 PHASE，我们设计了一个社会识别任务和一个社会预测任务。 PHASE 通过人体实验得到验证，证明人类感知到社交事件中的丰富交互，并且模拟代理的行为与人类相似。PHASE is a physics-based dataset for abstract social events. It incorporates social concepts such as helping another agent, and mirrors a broad range of real-world social interactions. PHASE comprises 2D animations of paired agents moving within continuous spaces, with all animations procedurally generated using a physics engine and a hierarchical planner. Agents have limited field of view, and can interact with multiple objects in environments containing multiple landmarks and obstacles. Using PHASE, we designed two tasks: a social recognition task and a social prediction task. PHASE has been validated via human subject experiments, which demonstrate that humans perceive rich interactions within these social events, and that the behaviors of the simulated agents align closely with those of human participants.

搜集汇总

背景与挑战

背景概述

PHASE是一个基于物理的抽象社交事件数据集，通过2D动画模拟代理在连续空间中的移动，程序化生成社交互动场景，用于社会识别和预测任务。该数据集经过人体实验验证，能有效反映人类社交行为。