Community Embedded Robotics: A Dataset to Study Perceived Social Intelligence and Safety During Unexpected Encounters with Quadrupedal Robots

<h1>
<h1>
Description
</h1>
<p>
This dataset derives from an interdisciplinary experiment designed to investigate perceived social intelligence (PSI) and perceived safety (PS) in the context of human-robot unexpected encounters in three scenarios motivated from the literature; (1) stop and back off: when the robot sees the person, it stops where it is before backing off and moving out of the participant’s way; (2) stop: when the robot sees the person it stops in place; and (3) efficiency: this behavior reflects many social navigation techniques that continues uninterrupted while treating the person as a dynamic obstacle to be avoided</p>

<p>The research questions that were investigated in this work were:</p>
<ol>
<li>How arePSI and PS related? Are higher levels of PSI associated with higher feelings of safety?</li>
<li>Are the statistical evaluations of PSI and PS representative of participants' feelings about these concepts during interview discussions?</li>
<li>How does the back-off behavior compare to the stop and efficiency behaviors with respect to PSI and PS during unanticipated crossings? (e.g. around corners, doorways, stairs).</li>
</ol>

<img src="https://dataverse.tdl.org/api/access/datafile/738535" alt="Time lapses of each of the three robot behaviors, shown in one of the three unexpected crossing scenarios.">

<p>The experiments investigated the three aforementioned human robot encounter scenarios and manipulated the autonomous robot behavior at the moment of those crossings. The experiments were performed in a wizard of oz manner, wherein a study member controls the robot from afar while the participant believes the robot is behaving autonomously. The behaviors tested were (1) the stop and back off (BO), (2) the stop in place (ST), and (3) to continue uninterrupted and treat the person as a dynamic obstacle, which we call efficiency (EFF). Time lapses of each of the three robot behaviors, shown in one of the three unexpected crossing scenarios are shown in the above figure</p>

<p>The study was run in two parts. First, we performed a (N=286) between-subject online video study, in which participants see a first person view video of a person in all three scenarios under one of the three behavior conditions. Along with demographic and personality characteristics, participants took a post-video survey of four Perceived Social Intelligence (PSI) Scales and Perceived Safety, taken from the Godspeed Questionnaire. The four PSI scales used are Social Competence, Identifies Humans, Rudeness, and Trustworthiness. </p>

<p>The second part of the study involved a (N=24) within-subject laboratory experiment where participants cross the robot in all three scenarios under each of the three robot behavior conditions. After encountering each behavior, the participants take the same survey used in the online video study. The purpose of this was to validate the low ecological validity, but high statistical power online video study results using a lower power, higher ecological validity laboratory study of crossings. In the laboratory study, participants were also interviewed to learn about their in depth perceptions and experiences during the robot encounters. Specifically, we interviewed them for 30-45 minutes following the experimental encounters and discussed PSI and PS and how they related to the robot’s behaviors. Particularly, the goal was to augment the statistical relationship between PSI and PS with a qualitative model that connects these two key factors used to evaluate perceptions of mobile robots. </p>

<p>The impact of this work and dataset is in the novelty of the human robot crossing scenarios investigated. Particularly, existing works in crossings in hallways, elevators or doorways are not designed to investigate unexpected crossings as studied here. Furthermore, this study utilizes a quadrupedal robot (Boston Dynamics Spot robot), which differs from many existing studies using various wheeled robots, including the Pepper robot or various autonomous robots.</p>

<p>This dataset provides all of the information necessary to enable the replication of the experiment The dataset can be used to study unexpected crossing scenarios between humans and robots at blind corners, blind doorways, and corners on stairwells. Particularly, the videos can be used to analyze path behavior and body language behavior from the participants in the experiments, which may offer valuable insights beyond the statistical and interview results presented in the accompanying paper. </p>

<p> Please, refer to the enclosed Data Report for more information about the data collection process </p>

<h1>Dataset Contents
</h1>

<h3>
Research Instruments
</h3>
<ol>
<li>Pre Experiment Questionnaire [PDF - 21KB]. This set of questions include Ten Item Personality Index and Demographic information about the participants and is taken before any robot stimulus.</li>
<li>Post Stimulus Questionnaire [PDF - 87KB]. This set of questions was taken after encountering the robot in each of the three experimental scenarios, under one of the three robot behaviors. Thus, it was taken one time by each participant in the online video study and three times by each participant in the real world laboratory experiment.</li>
<li>Semi Structured Interview Protocol [PDF - 104KB]. This was a list of questions that the research members used in interviews with participants in the real world laboratory study. They enabled the investigation into the perceptions and feelings of safety and social intelligence during the experimental encounters.</li>
</ol>

<h3>
Online Video Study Dataset
</h3>
<ol>
<li>Raw Survey Data [CSV - 100KB]. Contains the responses to the pre-experiment questionnaire and post-stimulus questionnaire in the online video study. Participant responses to the pre-experiment questionnaire are given in rows A-N for all participants. Participant responses to the post-stimulus questionnaire are given in columns O-Z or AA-AL or AM-AX, depending on which robot behavior was seen.</li>
<li>Robot Videos: Back-off robot behavior [MP4 - 8.6MB], Efficiency robot behavior [MP4 - 6.4MB], Stop robot behavior [MP4 - 7.4MB]. These are the video stimuli used in the online study. Each video shows the Boston Dynamics Spot performing one of the three behaviors in each of the scenarios.</li>
</ol>

<h3>
Lab Experiment Study Dataset
</h3>
<ol>
<li>Raw Survey Data [CSV - 13.5KB]. Contains the responses to the pre-experiment questionnaire and post-stimulus questionnaire in the online video study. Participant responses to the pre-experiment questionnaire are given in rows A-N for all participants. Notably, Column C of the document shows the order of behaviors seen by participants while Row 30, Columns B-D explain the meaning of the values in Column C for each participant. Columns Q-AB were the responses to the first behavior stimulus; Columns AC-AN were the responses to the second behavior; Columns AQ-AZ were the responses to the final behavior.</li>
<li>Human Robot Videos. This folder contains a sub-folder for each of the 24 participants. Under each participant are 9 videos, which are de-identified videos of each of the 9 crossing scenarios between the human and the Boston Dynamics Spot.</li>
</ol>

## 数据集说明
本数据集源自一项跨学科实验，旨在探究三类源自文献的人-机器人意外相遇场景下的**感知社会智能（Perceived Social Intelligence, PSI）**与**感知安全（Perceived Safety, PS）**。三类场景具体如下：
1. 停下后退：当机器人发现人类时，先停留在原地，随后后退并为参与者让出通行路径；
2. 原地停下：当机器人发现人类时，直接停留在原地；
3. 高效通行：该行为模式对应诸多社会导航技术，即机器人将人类视为需避让的动态障碍物，保持通行状态不中断。

本实验旨在探究以下研究问题：
1. 感知社会智能与感知安全存在何种关联？更高水平的感知社会智能是否与更强的安全感知正相关？
2. 针对感知社会智能与感知安全的统计评估，是否能反映参与者在访谈讨论中对这两类概念的真实感受？
3. 在意外横穿场景（如拐角、门口（doorway）、楼梯区域）中，后退行为与停下、高效通行行为在感知社会智能与感知安全维度上的表现有何差异？

上图展示了三类机器人行为在其中一类意外横穿场景下的延时摄影画面。

实验针对上述三类人-机器人相遇场景展开，在横穿瞬间对自主机器人的行为进行操控。实验采用**绿野仙踪范式（Wizard of Oz）**开展：研究人员远程操控机器人，而参与者误以为机器人处于自主运行状态。本次测试的三类机器人行为分别为：
1. 停下后退（Back-off, BO）；
2. 原地停下（Stop in place, ST）；
3. 保持通行并将人类视为动态障碍物的高效通行模式（Efficiency, EFF）。

本研究分为两个阶段：
第一阶段为**被试间设计的线上视频实验**，共招募286名参与者。参与者将观看三类场景下分别对应三种机器人行为的第一视角视频。除收集人口统计学特征与人格特质数据外，参与者还需在观看视频后完成四份感知社会智能量表与感知安全问卷，该问卷改编自**GODSPEED量表（Godspeed Questionnaire）**。本次使用的四份感知社会智能量表分别为：社会胜任力、识别人类能力、粗鲁程度与可信度。

第二阶段为**被试内设计的实验室实验**，共招募24名参与者。参与者需在三类机器人行为条件下，完成三类场景的横穿实验。每完成一种机器人行为的相遇后，参与者需填写与线上实验完全一致的问卷。本阶段实验旨在验证线上视频实验的结果：线上实验虽统计效力较高，但生态效度较低，而本次实验室实验虽统计效力较低，但生态效度更高，以此验证前者的结论。

在实验室实验中，研究人员还对参与者进行了访谈，以深入了解其在机器人相遇过程中的感知与体验。具体而言，实验结束后，研究人员将与参与者进行30至45分钟的访谈，讨论感知社会智能、感知安全及其与机器人行为的关联。本研究的核心目标之一，是构建定性模型，以补充感知社会智能与感知安全之间的统计关联——这两类因素均为移动机器人感知评价的关键指标。

本数据集与相关研究的创新性在于其所探究的人-机器人横穿场景。目前已有的横穿场景研究多聚焦于走廊、电梯或门口等环境，并未针对本文所研究的意外横穿场景展开。此外，本研究使用的平台为**四足机器人（quadrupedal robot）**——**波士顿动力Spot机器人（Boston Dynamics Spot robot）**，这与多数使用轮式机器人（如Pepper机器人或各类自主移动机器人）的现有研究有所不同。

本数据集包含复现本实验所需的全部信息，可用于研究人类与机器人在盲拐角、盲门口以及楼梯间转角处的意外横穿场景。具体而言，实验视频可用于分析参与者的路径行为与肢体语言行为，这或将为伴随论文中呈现的统计与访谈结果提供额外的有价值见解。

请参阅随附的数据报告，以了解更多关于数据收集流程的信息。

## 数据集内容
### 研究工具
1. 实验前问卷 [PDF - 21KB]：该问卷包含十项人格指数量表（Ten Item Personality Index）与参与者人口统计学信息，用于在机器人刺激实验前收集数据。
2. 刺激后问卷 [PDF - 87KB]：该问卷在参与者完成每一类机器人行为对应的实验场景后填写。因此，线上视频实验的每名参与者仅需填写一次该问卷，而真实世界实验室实验的每名参与者需填写三次。
3. 半结构化访谈提纲 [PDF - 104KB]：该提纲为研究人员在实验室实验访谈中使用的问题清单，用于探究参与者在实验过程中的感知与安全、社会智能相关的感受。

### 线上视频实验数据集
1. 原始调研数据 [CSV - 100KB]：包含线上视频实验中实验前问卷与刺激后问卷的所有回复。所有参与者的实验前问卷回复以行A至行N呈现。刺激后问卷的回复则根据参与者观看的机器人行为类型，分别以列O至列Z、列AA至列AL或列AM至列AX呈现。
2. 机器人行为视频：后退行为视频 [MP4 - 8.6MB]、高效通行行为视频 [MP4 - 6.4MB]、原地停下行为视频 [MP4 - 7.4MB]。上述视频为线上实验使用的刺激素材，每段视频均展示了波士顿动力Spot机器人在各类场景下执行对应行为的过程。

### 实验室实验数据集
1. 原始调研数据 [CSV - 13.5KB]：包含实验室实验中实验前问卷与刺激后问卷的所有回复。所有参与者的实验前问卷回复以行A至行N呈现。值得注意的是，文档的列C展示了每名参与者所体验的机器人行为顺序，而行30的列B至列D则解释了列C中各数值的含义。列Q至列AB为参与者对第一种机器人行为刺激的回复；列AC至列AN为对第二种行为的回复；列AQ至列AZ为对最后一种行为的回复。
2. 人-机器人交互视频：该文件夹包含24名参与者各自对应的子文件夹。每个子文件夹下包含9段视频，均为经过匿名化处理的、人类与波士顿动力Spot机器人在9类横穿场景下的交互视频。