IROS2019国际学术会议论文集1564

Right of Way, Assertiveness and Social Recognition in Human-Robot Doorway Interaction Jack Thomas1and Richard Vaughan1 AbstractWe expand on previous work for negotiating human-robot navigation contention around doorways to pro- duce a more socially-compliant autonomous robot behaviour. Our goal is to improve the integration of robots navigating in human environments by eliciting human recognition of the robots right of way. This is achieved by incorporating feedback from a user study of our previous system to create a more communicative, reciprocal, and assertive behaviour. Our contribution includes both the updated behaviour and a new user study that evaluates and compares the system to its predecessor. Results show that participants are more likely to respect the robots right of way given the new robot behaviour, but their responses also highlight the challenges of socially integrating robots into human spaces. I. INTRODUCTION Major corporations are poised to usher in a new generation of autonomous robots that share space with humans. The self-driving car is the most prominent example, but every- thing from patrolling security robots to fl ying delivery drones to friendly advertising humanoids is under development. Bystanders may one day discover their streets, skies, and workplaces fi lling up with autonomous machines - and the success of these ventures could hinge on the publics willingness and ability to interact with these robots. To explore the integration of robots into human envi- ronments, we previously proposed a method for handling navigational contention around doorways 1. A deadlock could occur when two parties on opposite sides of a doorway both want to cross to the other side, so the symmetry must be broken somehow. Human social etiquette has strong norms that robustly and effi ciently allocate the right of way. Previously we demonstrated an assertive behaviour could resolve these deadlocks, but this was only effi cient, i.e. resolved in shortest time, when the user was willing to acknowledge the robots right of way when appropriate. Our previous robot did not maintain any belief about who had right of way until after a deadlock was detected, thus no such belief could be signalled to the user to avoid the deadlock occurring. Deciding the right of way in advance was left up to the user. In this paper the robot maintains an early estimate of who will have right of way should contention occur, and modifi es its behaviour and appearance to signal this to the user. This issue of reciprocation and the split amongst partici- pants between those who would and wouldnt defer provides a case study for achieving recognition for autonomous robots. 1Jack Thomas and Richard Vaughan are with the School of Computing Science, Simon Fraser University, 8888 University Drive, Burnaby, B.C., Canadajacktsfu.ca vaughansfu.ca Fig. 1.A robot deferring to a human, a human deferring to a robot, and a deadlock caused by neither deferring. By establishing that the effi ciency of the interaction depends on human perception, we can measure how more socially nuanced behaviour impacts the performance of the system. This has motivated a redesign of the assertive behaviour according to feedback produced by the previous study. The resulting assured system aims to communicate its intentions more clearly, deepen its sensitivity to human movements, and - most signifi cantly - take a proactive role in asserting its right of way. To complete the cycle, the new system was evaluated in a second user study, functioning as a test both for our behaviour and this approach to social human-robot interaction development. The contributions of this paper are (i) an improved doorway-navigating behaviour for autonomous robots in hu- man environments, (ii) a study evaluating its performance with untrained users, and (iii) a cross-study analysis of pursuing robot social recognition as a means of integrating human-robot spaces. II. RELATED WORK A. Autonomous Navigation in Human Environments Navigation has been understood as a fundamental compo- nent of robot autonomy since the early days of the fi eld, and navigating among other autonomous agents as a key stepping stone toward working with humans. The Velocity Obstacle 2 is a popular and rigorous framework for au- tonomous robots to model other agents movements in their vicinity and formulate a motion plan to avoid collisions. Reciprocal velocity obstacles (RVO) 3 extended the idea by acknowledging that if other agents were also modelling each others behaviour, they could cooperate toward more effi cient outcomes. Integrating robots with humans takes more than just a multi-agent approach, however. Guzzi et al. 4 weighed RVO models against pedestrian models designed for human- like behaviour. In simulation, the human-like models outper- 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) Macau, China, November 4-8, 2019 978-1-7281-4003-2/19/$31.00 2019 IEEE333 formed the RVO-based ones by adopting many of the same lane and passing behaviours that emerge with large crowds. This demand for additional layers of awareness and sen- sitivity has led to many models of human social behaviour. The study of proxemics, for example, examines what socially acceptable distances should be maintained among humans, such as with Mumm and Mutlu 5, where a robots likeabil- ity and gaze behaviour was found to impact the distances humans were comfortable maintaining from them. Recently machine learning has been leveraged to infer human social preferences from data. In Everett et al. 6, for example, a deep reinforcement learning approach toward natural pedestrian behaviour through interaction is proposed to avoid the possibly ungrounded assumptions and approxi- mations that may hinder model-based approaches. B. Social Human-Robot Interaction Several authors have articulated the importance of a robots social agency to its autonomy in human spaces. Scassellati and Breazeal 7 argued that humans need to think robots have beliefs and intentions of their own to sustain a social interaction with them. Breazeal would go further 8, arguing for developers to adopt the robots view entirely as an equal partner to the human. This paradigm is more than abstract philosophy, it can inform practical development decisions. In Khambita et al. 9, motion planning for robots around humans is conceived of as a joint-action problem, where the robot considers the well-being and success of both parties. For their part, regular people are already primed to read meaning and intent into the actions of robots. Saerbeck and Bartneck 10 found users could read different affects into robot movements while varying only acceleration and trajectory curvature, while Mutlu et al. 11 found that slight, seemingly-unintentional gaze cues were enough of a hint from humanoid robots for humans to guess the answer in a hidden-object guessing game. C. Socially-Compliant Navigation Other researchers have made recent attempts to respond to many of the coming human environmental pressures we identifi ed, leading to navigation projects seeking compatibil- ity with humans. Burgard et al. 12 presented a planning method sensitive to the nuances of social distances for humans in motion, though they specifi cally imagine the robot as a subordinate rather than equal navigator - they explain their aim as “a polite and obedient robot navigation behavior that gives priority to humans and includes making way for them.”. Kandal et al. 13 describe another method for robots to socially coordinate spatial resources with humans, where robots recognize and acknowledge the superiority of their human interlocutors. While these projects explore the potential of robots as social navigation agents in human environments, studying the robot as a fully integrated equal remains underexplored. Fig. 2.Outline of the Previous Assertive System for Negotiated Doorway Interaction Fig. 3.Outline of the New Assured System for Negotiated Doorway Interaction III. SYSTEM Our assertive robot behaviour 1 is a means to resolve situations where a robot wants to pass through a doorway when a human on the far side wishes to do likewise. Here we iterate on this behaviour to provoke greater cooperation from humans, as part of a long term project of making robots that are easy to be around due to non-linguistic social compliance. A. The Assertive Behaviour The prior system was inspired by another, older method for resolving similar navigation deadlocks in corridors between autonomous robots, known as the aggressive approach 14. Under that method, two robots traveling down opposite ends of a narrow corridor would resolve their navigation deadlock in a fi ght. Both robots backed away from each other until the more aggressive one (defi ned as “further from their starting point”) would switch to advancing and push the retreating robot out of the corridor. Informal testing at the time suggested the same behaviour might be effective in human-robot interaction. This idea was developed into the assertive behaviour. This new behaviour was designed to be agnostic to the identity of the subject the robot was interacting with, whether human or robot. It replaced the fi ghting mechanism by having the robot stop in- place and wait to see if the other party would either advance toward them or retreat, before responding in kind. The state graph for the old assertive behaviour is shown in Figure 2. The previous study evaluated the assertive behaviour with three hypotheses: 1) That it would resolve doorway deadlocks, which it did in almost all of the studys 200 interactions. 2) That performance would be improved if people respected the robots right of way, which was clearly established when trials where the robot was not respected took longer to complete. 3) Respecting the robots right of 334 way would correlate with recognizing the doorway interac- tion as social, which qualitative evidence from questionnaires and surveys gave reason to believe. The studys results may have supported our theories con- cerning the role of social recognition, but actual acceptance of the behaviour was mixed. Participants deferred to the robots right of way in roughly half of the interactions where they were free to react naturally, with a quarter of participants never cooperating and another quarter always cooperating. The desire to increase that share and learn more about what factors were infl uencing this split in participant behaviour has motivated another round of development. B. The Assured Behaviour Reviewing participant feedback drew attention to a few issues, notably clear communication of intent and picking up on smaller signals the participants were putting out in return. These suggestions coalesced into the idea of proactivity - that the robot should take a more active role in resolving the deadlock and signalling its intentions, rather than waiting for the other party to act. This was achieved by three modifi cations: 1) State Feedback: The previous study made limited use of feedback to avoid distracting from the performance of the navigation behaviour, but many participants requested visual and audio cues to signal the robots intentions. Through the use of an LED light strip and digital speakers, the robot now indicates state changes by changing light colours and chirping. 2) Awareness while Deferring: If the robot plans to defer to the subject, it begins pulling over well before the door itself to provide space to pass, sending a clear signal that invites their interlocutor to cross through the doorway confi dent that the robot will not block their path. The robot also takes note of whether the subject approaches more from their left or right, and makes sure to defer in the opposite direction to avoid awkward collisions. 3) Assertiveness While Advancing: Rather than waiting to see what the subject does and then reacting, the robot now predicts right of way based on each partys respective distance to the door. Additionally, if the robot expects itself to have right of way, then it begins accelerating to clear the door faster. This presumes the cooperation of their interlocutor in resolving the situation smoothly, while also signalling their awareness of the subject and intentions regarding the door. In this study we test this as a single new system. We do not attempt to measure the contribution of each component, as each is non meaningful alone. Figure 3 describes the high-level states for the new assured behaviour, named for the confi dence with which it now asserts its right of way. IV. STUDY Having identifi ed a link between performance and social recognition with the last study, the next step is leveraging that link. This motivation was formalized into three new hypotheses: H1) More participants will respect the robots right of way with our new, feedback-informed behaviour. H2) More participants will recognize their interaction with the robot as social. H3) Pursuing robot social recognition through rounds of studies and development is an effective means to achieve integration of human-robot environments. The new study was necessarily structured similarly to the previous one, to allow for comparison between the the two. A. System Implementation We implemented the assured behaviour as a ROS1system, using open-source packages to handle navigation, mapping, hardware and sensor management. The robot used was a Pioneer-3DX, with a 270-degree hokuyo laser range-fi nder for obstacle avoidance, a Dell XPS laptop for onboard computation, and a top speed of 0.5m/s (which increases to 0.8m/s when asserting its right of way). The same robot platform and speed limit as the previous study were chosen to enable comparison of results, while off-the-shelf packages were used so that our approach would not depend on special, custom-built supporting software or hardware to function. Conversely, some questions of feasibility answered in the previous study led to streamlining in our new imple- mentation. Rather than use online mapping and a hand- crafted subject-detector for localization, a Vicon motion- tracking system was used to track markers on the robot and a helmet worn by participants. Similarly, having established the compatibility of the previous behaviour with purely robot-to- robot interactions, another round of all-robot testing was not explored. B. Environment The study took place in an 8m x 10m lab environment with a modular wall and open doorway constructed to block off one third of the space. This smaller space contained a desk for the test conductor and participant positioned 2m from the door, while the larger space contained a table with an open box positioned 4m from the door. C. Study Procedure Each test was split into fi ve trials. For each trial, the participant and test conductor would start at the desk, where the participant would be instructed to deliver some paper- work to the box on the far side of the lab. The robot would simultaneously be brought from its waiting position by the box to the desk, ensuring a doorway interaction with the participant. Once both had arrived at their destinations, the participant would then be recalled to the desk and the robot sent back to its station, resulting in a second interaction. The varying distances of the box and the desk from the doorway created a natural right of way for whomever started each interaction by the desk. 20 participants were recruited from the universitys student population, 10 men and 10 women, who were not com- pensated for their participation. The universitys offi ce of 1 335 Fig. 4.A study participant defers to the robots right of way. research ethics approved the project. The fi ve trials were unchanged from the previous study: 1) First Reaction Trial: Without being informed as to the details of the experiment, the participant was asked to drop their signed consent form in a box on a table in the larger room, while the robot would be called in to collect the reusable part of the form. The participant apparently incidentally interacted with the robot around the door as a result, and then again on their return journey. The participant was then informed that these incidental interactions were actually one of the trials, and the intent of the study was to examine their interactions with the robot around the doorway. 2) Teleoperation Trial: The participant was informed that the test conductor will take direct control of the robot via a controller, but to otherwise focus on the robot when deciding when to pass. The test conductor does their best to navigate the robot through the interaction via teleoperation, without being constrained to defi ned behavioral rules. 3) Full System Trial: The participant was informed the full autonomous system would now be active and the test conductor would no longer be in control of the robot. 4) Directed Behavior Trial: For the fi rst interaction, the participant was instructed to treat themselves as having absolute priority over the robot, and that the robot should defer to them. For the second, they were told to now treat the robot as having full priority, and that they should defer to it. 5) Full Explanation Trial: Before beginning the trial, the test conductor fully explained how the system worked to the participant, without instructing them on whether or not to obey it. As with the previous study, the fi ve trial were completed in the given order. The fi rst and fi fth trial, requiring fi rst impressions and acclimation respectively, could not be varied in their ordering, while the fourth trial depended only on participants knowing what would be expected of them to carry out their directions. It may have been possible to vary the second and third trials, but as these were not varied in the previous study, they were left in place. After each trial, the participant was given a survey to comple