Dancing with Robots at a Science Museum: Coherent Motions Got More People To Dance, Incoherent Sends Weaker Signal

Much has been said about the power of in-groups in social robotics. For example, people may favor a robot on their own team over a human on another team during game-play. Will this effect also hold over in a science museum exhibit illustrating interactive robot motions? Results from our study at the Oregon Museum of Science and Industry with 17 in-the-wild participant groups showed that coherently moving robots had the strongest communication signals, irrespective of in-group, and that incoherent motion can cue more complex communications. This paper presents two studies using a multi-robot expressive motion feature set called MoTiS. MoTiS consists of six parameters: direction, coherence, speed, start time, proximity and geometry. The first study explores coherent and incoherent direction from two different human perspectives: interaction partner and bystander, finding expressive features to be invariant to participant observation position. Our second study was in the wild at a museum, in which (1) participants were told they were assigned to a group - either Heroes or Villains, (2) some number of robots would greet participants (both groups or single group), and (3) some number of robots would lead them to a dance floor and dance with them. Irrespective of in-group, people were most likely to dance with robots when all the robots behaved socially, and when they were assigned to be affiliated with the Hero robots.


INTRODUCTION
This work acts as a finale study in a four-year development of a multi-robot motion control system called MoTiS [5], specifically focusing a parameter we call coherence, which we define for the purposes of this paper as multi-robot actions that occur with the same motion characteristics, e.g., all moving the same direction at the same time with similar motion onset and offset times, and contrasts to incoherence, in which robots might have have contrasting spatial or temporal motion characteristics.
For example, if a group of humans approached you directly and quickly, you might infer that they were looking for you, or worry you were the one that missed the memo, and should have already found them first.Depending on the context, like if they were already your friends, and depending on how they collectively approach, you may feel positive (e.g., like a group welcoming you), neutral (e.g., group of delivery people pulling a large package), or negative (e.g., a group of rowdy teenagers coming to rob you, yikes!).
Research has found homogeneous robot groups can be intimidating [14], i.e., when the robots looked similar and did the same thing, participants rated their behavior as threatening.While this prior study occurred in a controlled environment, we wondered how perceptions of robot group motions might vary in the context of two robot groups.What is the power of homogeneous moving together in this case?Will that vary with human affiliation?
To answer these questions, we begin with a baseline study looking broadly at coherence as a concept via an in-lab in-person user study, and expand on these results with a museum study involving a socially-communicative multi-robot interactive demonstration.The second study brought the robots into the wild, adding the concepts of affiliation and coherent/contrasting motion.Specifically, a group of hero robots and a group of villain robots that tried to attract museum goers to their respective dance floors, one group or both together.The results validate the communicative power of robot motion coherence in diverse in-person settings.In particular, subgroup size linearly affects communication strength, with larger subgroups having stronger communication.
As multi robot groups are becoming prevalent in human spaces, we have illustrated coherence as an increasingly relevant social signal to designing group motion and group behavior.The social psychology of meeting one versus several is clear.While coherence in a robot group may be intimidating to passersby in some cases, they might also not want to be so approachable that they invite interaction when they have a task to fulfill, thus moving together could clarify what subgroups/individuals robots are busy vs. available for service.

RELATED WORK
To understand how relative direction, level of coherence, and subgroup size can influence how people perceive multi-robot groups, we look to prior work in relative direction in expressive motion, human group dynamics, and coherence in multi-robot groups.

Expressive Direction and Trajectory
Relative direction has been explored for robot groups in work by Berger et al. [7].Three robots were placed in front of a human figurine and performed different motions at two speeds.Participants rated these motions in an online study.Fast motion towards the figurine was seen as aggressive and confrontational, but slow motion was seen as welcoming and excited.Away and sideways motion were seen as fearing the figurine at both speeds.
Trajectory has been explored in single robots and multi-robot groups.People view single-robot direct motion as goal-oriented and confident [27,28], while indirect motion is viewed as hesitant and confused [25,28].Similar results in multi-robot groups showed that trajectory of a multi-robot group affected a users ability to guess a group's goal [8].

Coherence and Groups in Human Behavior
People determine social groups through similar behaviors, culture, appearance, and identity [2,40].Humans within social groups often think in terms of in-group and out-group, which can often lead to a negative impression of perceived out-groups [2].These larger social groups often contain smaller, independent social groups, which is referred to as clustering [35].Clustering also occurs when subgroups have different functional roles within a larger group [2].
Coherent groups of humans are seen as powerful and influential, especially if large, and often have negative connotations, such as mobs and protests [44,46].The larger the group, the stronger their communication, influence, and in-group conformity [44,46].People are more likely to be influenced by information coming from majority groups than minority groups, [17] and group size directly affects how influential the group is [23].These groups create pressure to conform to the group's standards, highlighting the communicatory and influential power that larger groups have over individuals and smaller groups [35].However, coherence in groups is not always viewed negatively.For example, sports teams behave coherently with a collective goal, but this is positively viewed as teamwork, rather than a threat to others [36].

Coherence in Multi-Robot Systems and HRI
Coherence can be applied to multi-robot systems.Prior work has explored when when people perceive multiple robots as a cohesive group, showing that temporally asynchronous groups were perceived as the most expressive and spatially synchronous groups were perceived as the most cohesive [33,42].Studies have found coherent robot groups to be perceived as threatening, intimidating, and exclusionary, especially when the act as out-groups relative to humans [14,41].This argues for the idea of robot ingroups.

Robot-Human In-Groups & Out-Groups
Prior work includes indications that people act more favorably towards groups of heterogeneous robots than homogeneous robots, theorizing that high entitativity of a homogeneous group is experienced as more threatening [14].Similarly, robots illustrating only in-group communication, and no communication to outside humans, led people to feel more negatively, whereas those including humans were more likely to be experienced favorably [15].This finding has been confirmed by other work citing multiple prior studies showing that people react negatively to robots who act as out-group relative to humans [41].However, when robots are part of a human's in-group, people are more likely to feel positively towards an in-group robot than an out-group human [16].Additionally, people are more likely to interact with groups of minimal robots than single robots [13].
Similar to in-group and out-group, robots have the potential to induce and portray feelings of inclusion and exclusion [11,34].One work showed that when a human and two robots played a ball-passing game, with the robots displaying three different behaviors: exclusion, inclusion, and over inclusion, the exclusion case caused the human to perceive that the robots were excluding them intentionally [11].
Work has also shown that motion can be a defining factor of subgroups within a multi-robot group.Robots that have the same relative direction are perceived as being part of a group, whereas robots not moving, or moving in opposing directions are viewed as individual robots or separate subgroups [4].
For this study, we utilized the relative motion and relative spacing, with our robot motion exploring variations on group direction and coherence within direction, and different geometrical formations.MoTiS has also been successfully used by non-experts to create custom, multi-robot expressive motion.In a prior workshop with non-robotics specialists, participants checked out and in code from its GitHub repository to create expressive pathways.Their sequences and robot application backstories utilized a bodystorm then program technique, as motion expressions are often clearest experientially.This workshop demonstrated that novice users, with the support of social robotics experts, could create novel multi robot expressive motion ideas using the MoTiS software.

Software Implementation of MoTiS
The base software is run using Robotic Operating System (ROS) and was implemented on the Pioneer 3DX robots.The system is fully distributed and does not include inter-robot communication.The MoTiS software implementation consisted of a local planner and a relative keyframe goals algorithm.First, a user inputs the relative goal, referred to as keyframe, for the robot and chooses which instances of parameters they would like to include, with +y being the forward direction of the robot and +x being the right direction of the robot.Not all parameters affect the final goal to the same extent.The choices for relative direction, geometry, and proximity are put into the relative keyframe algorithm, and the choices for speed and start time go into the local planner.After the user specifies which instances of the parameters they want the robot to run, the keyframe (and in some cases relative speed) is sent to the local planner.The local planner is an algorithm that takes in the sensor input and current goal and outputs the forward and rotational speed to the pioneer controller package.Sensor data is received from the rosaria and pioneer_sensors packages, and the output of the local planner is sent to the low level controller package, pioneer_controller.We refer to this as the keyframe approach.

IN-LAB STUDY: INTERACTION PARTNERS VS BYSTANDERS
Our in-lab study explored how humans perceive and react to four Pioneer 3DX robots moving coherently and incoherently towards and away from them from two perspectives: bystander and interaction partner.We aimed to to see if results would be the same for someone directly interacting with the robots vs someone observing the interaction.We also aimed to test if our results from our prior work using online studies to explore coherence [4] translated to inperson human-robot interaction.There were 12 total participants in 6 groups.

In-lab Study: Study Conditions
For our lab-run coherence study, we replicated the explicitly incoherent study conditions from the online study on four Pioneer 3DX robots [4], seen in Fig. 3.In addition to the coherence conditions, we added two human condition, which were the roles each participant (N=12) played in the interaction with the robots: (1) interaction partner and (2) bystander.As seen in Fig. 3, the interaction partner stood centered, facing the robots, similar to the role of the figurine in Studies 1 and 2. The bystander stood off to the right of the interaction partner and observed the interaction between the robots and the interaction partner from afar.

In-lab Study: Study Procedure and Evaluation Methods
For this study, we ran two participants at a time, with one participant as the interaction partner, and one as bystander.Both participants were given a consent form.Once signed, the participants would go through all five motion conditions in a random order.Participants would be randomly assigned either the interaction partner or the bystander.Once assigned, the participants would get into position and the robots would execute their motion condition.Participants would then be lead to a back table where they would fill out a survey, with questions seen in Table 1.After all five conditions, participants had a short debriefing interview to relate any notable experiences during the experiment.

In-lab Study: Hypotheses
It was hypothesized that coherent motion will have the strongest communication signals (H1.1) and that there will be a linear relationship between subgroup size and communication strength (H1.2).It was also hypothesized that coherent motion towards will be the most threatening, inviting, and blocking (H2.1) and coherent motion away will be the most harmless, avoiding, and not blocking (H2.2).These hypotheses were based on the results of our prior online study [4].Finally, it is hypothesized that bystanders will perceive the robots as overall more neutral than interaction partners (H3).
Human Position Results.Against our hypotheses, there was no significant differences between how interaction partners vs bystanders ranked attributions.
Threatening / Harmless.Coherent motion towards was viewed as the most threatening by bystanders and in the combined data supporting hypothesis H2.1.Coherent motion away was viewed as the most harmless for interaction partners and in the combined data supporting hypothesis H2.2.Generally, there was a linear tendency, with coherent motion in both directions being viewed as the most extreme, and each condition being viewed as more harmless the larger the away subgroup became.There was no significant difference between the results of the interaction partners and the bystanders, not supporting hypothesis H3.Results shown in Fig. 4.
Avoiding / Inviting.Coherent motion towards was seen as the most inviting for interaction partners and bystanders, supporting hypothesis H2.1.Coherent motion away was also seen as the most avoiding, significantly more than all other study conditions in the combined data, supporting hypothesis H2.2.Again, there was a linear tendency in the combined data, with communication strength increasing with subgroup size, supporting hypotheses H1.1 and H1.2.There was no significant difference between the results of the  interaction partners and the bystanders, not supporting hypothesis H3.Results shown in Fig. 5 Blocking / Not Blocking.Both coherent motion towards and 3 Towards, 1 Away were viewed as very blocking, however results were not significant, not supporting hypothesis H2.1.Coherent motion away was overall seen as the least blocking, supporting hypotheses H2.2.However, bystanders rated 1 Towards, 3 Away as less blocking than coherent motion away. 2 Towards, 2 Away was rated neutrally between both interaction partners and bystanders, creating a linear tendency between subgroup size and communication strength.These were not statistically significant, not supporting hypotheses H1.1 and H1.2.There were trends ( < 0.1) showing that interaction partners rated groups with the majority moving towards as more blocking than bystanders, somewhat supporting hypothesis H3.Results shown in Fig. 6.

In-lab Study: Discussion
Coherent motion led to strong communicatory signals (Figs. 4,5,6).These results show coherent motion gave the clearest communicatory signals, while incoherent motion had large variances and lacked significant differences between the conditions in many cases.The communicatory strength of coherent motion may be due to the fact that all of the robots are exhibiting the same behavior.This phenomenon also appears in humans groups [17,35,37,44].In these human situations, groups exhibiting the same behavior send very strong communicatory signals.Our results expand the findings of work in human behavior to robot groups [17,35,44].
Coherent motion, especially towards, was perceived more negatively than incoherent motion, (Figs.4,5,6).Coherent motion towards was the most threatening and blocking.Coherent motion away was the most avoidant.Directional coherent motion may lead participants to associate the robot group with crowd behavior associated with negative emotions, such as protests [20,37] and evacuations [22], therefore perceiving the robots as aggressive or scared.This interpretation was supported by quotes from participants in our in-lab coherence study.One participant said coherence towards that the robots were "trying to communicate a hostile feeling." Multiple participants said of coherence away that the robots reminded them of "scared animals, " or "fleeing." Subgroup size linearly affected communication strength (Figs.4,5,6).Subgroups of three/one had stronger signals than subgroups of two.This was most apparent in the combined interaction partner and bystander data.While the interaction partner data shows consistent linear trends, the bystander data was more ambiguous, which may be because the bystander was merely watching the interaction and did not perceive the robots from the same perspective as the interaction partner.In human situations larger groups yield stronger communication [17,35], and this effect may extend to robot groups.
Interaction Partners and Bystanders tell different stories.There were often differences in the stories interaction partners and bystanders told about what they thought the robots were doing and what they reminded them of.In one case of coherent motion towards, the interaction partner thought that the robots were coming towards them "maybe [out of] curiosity, " while the bystander saw four robots beeline towards the interaction partner and thought the interaction partner was "their leader and they are trying to show deference to her...[to] attack!"Sometimes, the interaction partner felt more threatened by the robots moving towards while the bystander was not.For example, the interaction partner thought the robots were "threaten[ing] me" while the bystander thought they wanted to "say hi."

IN-THE-WILD HEROES VS VILLAINS STUDY
This study brought the robots into a public museum installation to utilize concepts of expressive motion in a dedicated human-robot context.The study took place at the Oregon Museum of Science and Industry (OMSI) at one of their After Dark Events: a 21+ evening event involving many booths with volunteers and local organizations, where people are able to walk around the museum, eat food, and explore all of the demonstrations at the different booths.This particular event was themed "Superheroes and Supervillains, " which we leveraged to explore the role of both group motion and affiliation.Upon arrival to our booth, participants were assigned as either heroes or villains, then the robots would then greet them, lead them to a dance floor, and dance.Over the course of the night, 28 participants participated in our study.

Heroes vs Villains: Study Conditions
There were two independent categories of manipulation: (1) motion and (2) affiliation.Motion and affiliation were designed to be part of the study.Motion described how the robots moved in the space and had two subcategories: Incoherent (Adversarial, M1) and Coherent (Everyone, M2).The affiliation described whether the participants were assigned to be affiliated with the Hero robots or the Villain robots.For the rest of this paper, we will use the terms "affiliate robots" and "non-affiliate robots, " which refers to the robots matching the assigned affiliation of the participant(s).For example, if a participant is assigned to be affiliated with the Hero robots, Hero robots are the affiliates and Villain robots are the non-affiliates.
The two group motion conditions included: Incoherent (Adversarial, M1) motion in which affiliate robots behave socially, moving towards the participant and dancing, while non-affiliate robots were anti-social, moving away from the participant and not dancing, and Coherent (Everyone, M2), in which all robots behaved socially, moving towards the participant and dancing.Both had three phases: Greet, Lead, and Dance (Fig. 1).In the Adversarial condition, the affiliate robots came towards the participants and did their gesture during the Greet phase, while the non affiliate robots move to guard their respective dance floor.During the Lead phase, the affiliate robots turn around and lead the participants to the affiliate dance floor and the non-affiliate robots continue guarding.Finally, in the Dance phase, the affiliate robots dance while the non-affiliate robots guard.In the Everyone condition all robots approach the participant, but only the affiliate robots greet with their gesture in the Greet phase.In the Lead phase all robots turn around and go to their respective dance floors.In the Dance phase all robots dance.Examples of both motion conditions are seen in Fig. 7.
The affiliation conditions were 1) Heroes (H) or 2) Villains (V).The motion design gestures were designed to match [26], which showed that front to back motion was seen as more aggressive and less polite than side to side motion.The villain gesture was chosen as front to back motion also because it is an angular motion, which has been shown to map to negative emotions in multi-robot groups [39].The hero gesture was chosen as a rotation from side to side, which prior work has shown to be interpreted as polite [26].This side to side rotation is also a round motion, which in multi-robot groups has been shown to map to positive emotions [39].

Heroes vs Villains: Methods
All the trials used the same 15ft by 15ft space inside of OMSI.The robots were laid out in a line in front of the participant starting location as shown in Fig. 8. Behind the robots lay their associated dance floors in corresponding colors.The participant starting location was next to a determination light that designated the participant to be a "hero" or "villain." Once a participant signed the consent form, they were guided by the first researcher to the determination light.The light turned either red for villain or green for hero.The color was predetermined for each trial and controlled by a second researcher standing off to the side.The robots' predetermined intended motion was then started by a third researcher outside the main interaction space.The motion consisted of a greet, lead, and dance.Participants were recorded during the interaction with two cameras: a GoPro shot from the side of the interaction space and a webcam suspended 10 feet above.Alongside the video data, comments made by the participant during the trial were recorded along with reactions.
We used two main methods for evaluating participant response to the robots.The first was subjective user reported surveys, which users filled out after their interaction with the robots.The surveys consisted of four anchored scale questions and four extended response questions.The anchored scale questions were on a seven point scale, with the negative adjective corresponding with a negative numerical rating and the positive adjective corresponding with a positive numerical rating.The four word pairs were threatening/harmless, avoiding/inviting, unwelcoming/welcoming, and unfriendly/friendly.After the four anchored scale questions, there were four extended response questions: 1) Did you attribute personality to the robots in this activity?If so, what sorts of personalities? 2) Did you feel like the robots wanted you to join them?Why or why not? 3)What do you think is the robots' motivation? and 4) Did the robots remind you of anything?The second was objective behavioral observation done by a study conductor.Behavioral observations primarily consisted of recording quotes from participants.Significant facial reactions (surprise, laughter, etc) along with motion (dancing, following, etc) were also recorded.One study conductor was designated to observe and note behaviors.
For the analysis of the motion conditions and affiliation conditions, only the data labeled as successful was used.This successful data across the anchored scaled questions was distributed normally when sorted by motion condition and by character affiliation, one way ANOVA tests were used to find pairwise significance for each question.For the behavioral measures, we annotated the data as binary if the paricipants followed or danced.This data was then tested using two-proportion z-test to see if there was significant differences across each independent variable pair (Adversarial vs Everyone motion, hero vs villain affiliation).

Heroes vs Villains: Hypotheses
It is hypothesized that Adversarial will be viewed as more avoiding (H1.2), more unwelcoming (H1.3), and more unfriendly (H1.4) than the Everyone condition.However, it is hypothesized that the Everyone condition will be viewed as more threatening (H1.1), as more robots are approaching and prior work has shown that multiple robots approaching together can read as threatening [4,5].Behaviorally, it was hypothesized that more people would follow the affiliate robots in the Adversarial condition (H1.5) because it would be more clear which group to follow.It was also hypothesized that there would be no difference in dancing across robot motion condition (H1.6).
It was hypothesized that people would rate the robots more negatively when they were assigned a Villain affiliation.Therefore, a Villain affiliation would be more threatening (H2.1), more avoiding (H2.2), more unwelcoming (H2.3), and more unfriendly (H2.4) than when affiliated with the Hero robots.For the behavioral measures, it was hypothesized that more people would follow their affiliate robots when assigned a Hero affiliation (H2.5) and more would dance when assigned a Hero affiliation (H2.6).

Heroes vs Villains: Results
Motion Conditions.Across the attribute ratings from the anchored scale questions, there was no statistical significance when comparing Adversarial to Everyone, which does not support hypotheses H1.1-1.4.Overall, participants viewed the robots positively across all attributes.Non-significantly, more participants rated Everyone as slightly less threatening (mean = 1.89) than Adversarial (mean = 2.38), not supporting hypothesis H1.1.More participants also rated Everyone as more welcoming (mean = 1.11) than Adversarial (mean = 0.75), however this was not significant, not supporting hypothesis H1.3.Participants were significantly (p = 0.02) more likely to dance with the robots during Everyone than during Adversarial supporting hypothesis H1.5 (Fig. 10).Non-significantly Both Adversarial and Everyone were described as "cute" and "curious" by eight different participants, which supports the numerical results showing that the robots were generally viewed positively regardless of movement condition.They were equally as likely to say the robots wanted them to join for both conditions.Adversarial was described explicitly as "threatening" by four participants, while Everyone was never described at threatening.This supports the numerical data where Adversarial was rated as more threatening.Adversarial was also described as "awkward" while Everyone was not.One participant said, "[the robot] wants to learn how to dance at a middle school dance." Another described the robots as "clingy, awkward." Everyone was also noted as being slightly more inviting than Adversarial, with 56% of participants saying that Everyone seemed inviting them to join, but only 43% of participants saying this of Adversarial.However, more participants described Adversarial as explicitly "friendly", with five calling Adversarial friendly and only two for Everyone.
Affiliation Conditions.There were no significant differences in the attribute ratings between H: Hero and V: Villain and these differences were even smaller than between motion conditions, not supporting hypotheses H2.1-2.4.There was a statistical trend (p = 0.09) that people were more likely to dance with the Hero robots than the Villain robots (Fig. 10), somewhat supporting hypothesis H2.6.More participants also followed the affiliated robots when they were assigned Hero affiliation than when they were assigned Villain affiliation (Fig. 9), but not significantly, not supporting hypothesis H2.5.Qualitatively, people were equally as likely to describe the robots as "friendly" or "cute" regardless of affiliation, which supports the numerical data showing that robots were viewed positively regardless of affiliation.Unexpectedly, people were more likely to say the robots wanted them to join when they were assigned a Villain affiliation.One participant said regarding the Villain robots "I'll follow you to your lair." However, people only thought the motivation behind the robots was to become friends when assigned a Hero affiliation.

Heroes vs Villains: Discussion
People were more likely to dance in the Everyone motion condition, which went against our initial expectation that motion variance would reinforce the social affiliation.One possible reason for this is that all the robots are dancing, and therefore the human feels more comfortable joining in, or felt peer pressure to dance since all the robots were dancing.It may also have been that participants in the Adversarial condition felt more uncomfortable due to the non-affiliate robots not moving and seemingly staring at the affiliate dance floor.This interpretation is supported by one participant who described the non-affiliate robots in Adversarial as being "judgy" and made them feel "rejected." On the other hand, more participants followed the affiliate robots during the Adversarial condition than the Coherent (Everyone condition).One potential reason for this may be that it is clearer which robots to follow when only the affiliate group approaches and leads.During the Everyone condition participants may feel like they have more of a choice of who to follow since both robot groups approach and lead back.This interpretation is supported by one participant who was assigned a Villain affiliation during the Everyone condition, but followed the Hero robots and then said, "I wonder if I'm supposed to follow the [villain]." Finally, human response varied across robot costuming and character (Hero and Villain).People were more likely to dance with the Hero robots, which may have been because they were seen as more friendly.This result may also have been due to the Hero character gesture being more similar to how people might dance in a small space: twisting back and forth instead of shuffling forwards and backwards.This interpretation is backed up by verbal comments participants made on both the Hero and Villain dancing gestures.One participant excitedly exclaimed "They're joining!" when there Hero robots began dancing, whereas one participant said skeptically of the Villains, "It's vaguely dancing, not humanoid dancing, but robot dancing." The character gestures may have played a part in more participants following the Hero robots than the Villain robots; the Hero gesture is less intrusive and may be friendlier.

DISCUSSION
We have demonstrated that group motion can be expressive and communicate intent and motivation to humans in the space.We saw in our in-lab coherence study that group motion may be more powerful than single robot motion, with an increase in communication strength proportional to the number of robots moving together.Complex expressive group motion is key in completing certain tasks, with coherence being a key part of this group motion.As seen in our in-lab study, coherent motion can send clear communication signals in terms of binary motivational descriptors, while incoherent motion can provide a more nuanced interpretations of group goals.We also saw evidence of this in our in-the-wild Heroes vs Villains study.People were more likely to interact and dance with the robots when the robots came up to greet them coherently.
We suggest that future researchers use fully coherent motion with caution, as it was seen to often have negative connotations.One example application which could benefit from fully coherent motion would be guarding an unsafe construction site.Here, a strong communication signal to stay away is needed due to safety hazards and a negative reaction from a person may actually help keep them away from the dangerous site.As seen in study 2, coherent motion can also be used in social situations to encourage certain behavior from a human in the space.However, incoherent motion is generally higher recommended for most social situations as it prompted more interpretations of the robot group's motivations.
We also saw between the two studies that interpretation of coherence can be context dependent.For example, in the first study coherent motion towards was seen as intimidating, but in the second study where the context was a dance party, coherent motion towards seemed welcoming.Future researchers should keep in mind social context when designing group motion.
Distributed systems are fairly robust to single robot failure, however expression is not.One or more robots failing, even though the other robots can continue, vastly changes the communicatory output of the group.For example, there was a case of the Coherent (Everyone) motion condition in which all robots approached the participants, but only the non-affiliate robots returned to the dance floor.This participant chose to follow the non-affiliate robots and dance with them instead, saying "Oh he's moving, so I'll go to him." This may have read as a rejection from the affiliate robots, since they approached the person and seemingly decided not to lead them back to their dance floor.
Despite best efforts, there will always be some cases in which robots in a multi-robot group fail.In a purely task-oriented multirobot group this is generally not catastrophic since the other robots can often replan and take over for the failed robot, complete failure of one or more robots in a group significantly impacts the expressive communication of the overall group.It is important to design expressive motion with the possibility of failure in mind, and design ways to get a similar communicatory effect in case of failure.The relative keyframe approach allows users to pick and choose which parameters they utilize, and allows them to specify different instances of those parameters which gives the user a high level of control over each desired waypoint.For example, the first relative goal may have one set of MoTiS parameters applied, and the second relative goal could have a completely different set of MoTiS parameters and instances.This allows users to decide where the next goal is in real time, which is beneficial in real world scenarios where the context and space can change.
Using a human-in-the-loop keyframe approach greatly contributed to the success of study 2. For example, there were times when one robot blocked another.Without the ability to move the blocking robot before it's final waypoint, the robots would have been stuck in a deadlock.For future researchers, this type of implementation is recommended so that some failure cases can be rectified.MoTiS makes this easier as future researchers can design keyframes with different aspects of MoTiS, such as "towards person" and can then combine different keyframes on different robots to change the geometry and coherence of the group.Additionally, it is important for researchers to explore which aspects of MoTiS are important to their communication goals and design keyframes around them.

CONCLUSION
In this work we first explored how coherent and incoherent motion (relative direction and subgroups) impacted people's perception of multi-robot groups.Results found that coherent motion, especially coherently towards and coherently moving away, send the clearest communicatory signals.Coherent-towards, in which all robots approach as once, is generally seen as aggressive, while coherentaway was seen as either avoidant and courteous.Coherent away motion was also sometimes seen as courteous and not blocking the human, and coherent stillness was the friendliest and most welcoming condition.On the other hand, while incoherent motion can be unclear in binary communication descriptors, it allows for more nuanced communication.Participants often described the subgroups as having different goals and motivations.Subgroups may also mimic human social behavior more closely, and therefore appear more social and less jarring than coherent motion.
In our next study, we included behavioral variants of coherence and incoherent robot group motions to a scenario of two teams of robots, heroes and villains, to see if the additional impact of affiliation would influence the previous results.Again we see that communication strength rises with number of robots, and only a slight influence of our affiliation variable.Moreover, this substudy acted as an in-the-wild demonstration of our MoTiS software and reliability, as live robot monitors could issue corrections should any individual robot miss its cue or go off path, enabling the installation experience to remain lively and clear.Both studies add new concepts from human group dynamics [2,17,35], such as subgroups and affiliation, that drawing on definitions of coherence in social psychology [1,2].Robot group communication insights, and these early findings about human behavioral responses to them can be utilized by future researchers -alongside the various other MoTiS motion features -to vary multi-robot group functional and social communications in shared human spaces.
Future researchers can utilize MoTiS to design and generate complex expressive multi-robot motion, allowing future researchers to more quickly program expressive motion for their own research areas.Using a common set of features could also aid in crosscomparisons and replicability in the small but growing efforts to include designed socially communicative motions into everyday robots.Future opportunities for extending this work include assessing MoTiS parameter variants across different social contexts and varied human-robot objectives.For example, one could generate legible multi-robot group motion, with emphasis on in-person studies with physical robots, such as bouncers for a club entrance, tending to an emergency, or guiding a group.
Compliance with Ethical Standards This work was conducted under IRB #8724.

Figure 3 :
Figure 3: In-Person study coherence conditions shown from most coherent (4 towards, 4 away), to least coherent (2 towards, 2 away).Human conditions of Interaction Partner (IP) and Bystander (B) are shown in each coherence condition.

Figure 4 :
Figure 4: In-lab Study: Coherent motion towards is threatening, while coherent motion away is harmless.

Figure 5 :
Figure 5: In-lab Study: Coherent motion towards is the most inviting, while coherent motion away is the most avoiding.

Figure 6 :
Figure 6: In-lab Study: Coherent motion towards is the most blocking, while coherent motion away is the least blocking.

Figure 7 :
Figure 7: Adversarial and Everyone motion conditions.

Figure 8 :
Figure 8: Starting setup from overhead for Heroes vs Villains

Table 1 :
In-Person User Study Questions Question Anchored Scale The robot group was [threatening / harmless].The robot group was [avoiding / inviting the human].The robot group was [not blocking / blocking the human].Extended Response What do you think the robot group was trying to do?What do you think is the motivation of the robot group?Did the robots remind you of anything?