Supportive Actions for Manipulation in Human-Robot Coworker Teams
Shray Bansal, Rhys Newbury, Wesley Chan, Akansel Cosgun, Aimee Allen, Dana Kulić, Tom Drummond, Charles Isbell
SSupportive Actions for Manipulation in Human-Robot Coworker Teams
Shray Bansal , Rhys Newbury , Wesley Chan , Akansel Cosgun , Aimee Allen ,Dana Kuli´c , Tom Drummond and Charles Isbell Abstract — The increasing presence of robots alongside hu-mans, such as in human-robot teams in manufacturing, givesrise to research questions about the kind of behaviors peopleprefer in their robot counterparts. We term actions that supportinteraction by reducing future interference with others as supportive robot actions and investigate their utility in a co-located manipulation scenario. We compare two robot modesin a shared table pick-and-place task: (1) Task-oriented: therobot only takes actions to further its own task objective and(2)
Supportive : the robot sometimes prefers supportive actions totask-oriented ones when they reduce future goal-conflicts. Ourexperiments in simulation, using a simplified human model,reveal that supportive actions reduce the interference betweenagents, especially in more difficult tasks, but also cause therobot to take longer to complete the task. We implemented thesemodes on a physical robot in a user study where a human and arobot perform object placement on a shared table. Our resultsshow that a supportive robot was perceived as a more favorablecoworker by the human and also reduced interference with thehuman in the more difficult of two scenarios. However, it alsotook longer to complete the task highlighting an interestingtrade-off between task-efficiency and human-preference thatneeds to be considered before designing robot behavior forclose-proximity manipulation scenarios.
I. I
NTRODUCTION
Despite the continued growth of industrial robot sales [1],many assembly tasks are still performed manually in majorindustries [2]. A vision for the future of manufacturinginvolves robots working alongside human coworkers on tasksthat exploit the respective strengths of both. Surveys identifyinteraction with co-workers as one of the most importantjob criteria for human workers [3]. We introduce interaction-supporting actions that aim to improve the coworker experi-ence in human-robot co-located manipulation. We implementthese in a close-proximity manipulation task to understandthe impact on task performance and the coworker perceptionas compared to a robot focused solely on completing its task.We term actions necessary for an agent to complete theirtask in the absence of other agents as task-oriented . We de-fine supportive actions as actions that support the interactionby reducing potential interference with other agents but arenot necessary for task completion. For e . g ., when resetting achessboard, for the agent playing black, actions that move theblack pieces to their positions are task-oriented and actionsmoving the white pieces towards the other player can be supportive . Although the supportive actions help the otheragent, they are not altruistic as the agent hopes to benefitfrom the reduced interference they cause. Humans also School of Interactive Computing, Georgia Institute of Technology,Atlanta, GA, 30308, USA. [email protected] Department of Electrical and Computer Systems Engineering, MonashUniversity, Clayton, VIC, Australia. Fig. 1. An example scene from our co-located manipulation scenario. Therobot’s goal is to place all the red blocks into the row closest to itself, andthe human participant’s goal is to do the same for the yellow blocks. perform supportive actions, perhaps due to their modelingof others as intentional agents that plan for mutual benefit[4], [5], or their expectations of reciprocity [6], etc .Our task is inspired by other close-proximity human-robotinteraction (HRI) manipulation studies [7], [8]. It involvestwo agents, a human and a robot, situated across a tablescattered with color-coded blocks, each aiming to bring theblocks of their assigned color quickly back to their side (Fig.1). The agents have intentionally been assigned separategoals without a direct incentive for cooperation and theshared table is expected to induce interference. We focuson high-level decision-making and design supportive actionsthat proactively avoid collision by modifying the goal con-figuration of the other agent by moving their blocks. Inour experiments, the robot operates in one of two modes:(1)
Task-oriented , where the robot takes only task-oriented actions, and (2)
Supportive mode where the robot takesboth supportive and task-oriented actions depending on thesituation. We hypothesize that the supportive mode wouldreduce interference and lead to a better human experience,both in terms of subjective and objective measures. We testthis in simulation with a simplified human model and verifyit with a user study on a physical robot.Our main contributions are the introduction of supportiveactions in a human-robot collaborative manipulation task,simulation experiments and user study experiments thatjustify the use of such actions, and the identification of atrade-off between operational and usability metrics when therobot is designed to deliberately take supportive actions.After reviewing the literature in Sec. II, we formulatethe problem in Sec. III and describe our methodology in a r X i v : . [ c s . R O ] M a y ec. IV. We first experiment in simulation in Sec. V to designthe supportive actions and formulate hypotheses in Sec. VI.We present the implementation and user study details inSections VIII and VII, respectively. We analyze results inSec. IX, discuss them in Sec. X and conclude in Sec. XI.II. R ELATED W ORK
Human-robot interaction (HRI) includes collaborative sce-narios where agents work to achieve a common goal andothers where agents have separate, sometimes competing,objectives. Our goal is to study scenarios where humans androbots work alongside each other, and interaction arises fromconflict due to shared resources (such as space).In manipulation, human-robot co-presence focus on sce-narios where the human is either treated as an obstacle tobe avoided [8], [9], or, as a leader [10] to be assisted. Inthe former, the human’s goal is either not considered at all,or only used to make predictions to guide a more pro-activeobstacle avoidance, and in the latter, the robot shares thehuman’s goal. Our task involves separate goals for the twoand we consider the question of whether the robot shouldtake actions that support the interaction without direct taskcompletion benefits for the robot.For assisting the human by anticipating their actions,Hawkins et al . [10] exploit task structure and Nikolaidis et al . [11] perform online adaptation to user preferences.Similarly to us, both of these approaches focus on thehigh-level decision-making aspect of the task. Cherubini etal . [12] plan low-level robot actions to successfully reducehuman workload for automotive manufacturing and Koppula et al . [13] perform assistive actions adapted to predictionsmade by learning model of the human’s activities.These methods have helped improve collaborative taskperformance, but the inherent assumption is that the roleassigned to the robot should be to assist and/or stay out ofthe way. These simplify the robot’s decision-making to favoractions that directly further the human’s objective. However,the types of roles and interaction modes in mixed human-robot teams are richer, as shown by Gombolay et al . [14].Similar to our task, Gabler et al . [7] plan robot actionsin a close-proximity human-robot collaborative scenario.Although both agents have a common goal in their task, theyutilize a game-theoretic model that considers the human asan agent with a different utility. They use this goal-drivenbehavior while planning to increase joint task-efficiency. Wedesign
Supportive actions to influence human behavior bymodifying goal configuration by moving their blocks. Thishelps to avoid future interference. While their model alsoconsiders the robot’s influence on the human, they only use itto find an optimal ordering of existing task-oriented actions.III. P
ROBLEM F ORMULATION
We design a pair of pick-and-place tasks on a table sharedby a human and a robot and represent it as a two-agent game.The table has two sets of blocks distinguished by color, weassign one set to the robot b R = { b R , ..., b nR } and the otherto the human b H = { b H , ..., b nH } . We draw a D grid on the
12 432 a a a a Fig. 2. An example board configuration consisting of blocks ( − )for the robot (red) and the human (yellow). Robot actions ( a − ) aredepicted by the arrows. a , a are task-oriented actions while a , a are supportive and a is a more useful supportive action because it reducespotential interference when reaching for block . table and place each block in a single cell. We define thiscell as the block’s location l ( b i ) = ( r, c ) and a cell near itsassigned agent as its destination, d ( b i ) = ( r, c ) . A state is aconfiguration of blocks on the grid, s = { b iR , b iH } ∀ i ≤ i ≥ n . Fig.2 shows a grid configuration where n = 2 and b R = { , } and b H = { , } .In this task, an action a can move at most one block toa different location. For instance, in Fig. 2, a = ( , d ( )) ,moves block from its location to the goal. We also allowidle actions that do not move any blocks. Both agents areinstructed to start performing actions simultaneously, and so,if one agent finishes their action early, they have to wait forthe other to complete their action before starting to performthe next one. We assign each agent the goal to take actionsthat lead to a state, s G , where each of their blocks is in itsdestination cell, in minimum time. Their goal only dependsupon their blocks, the locations of other blocks is not directlyrelevant. IV. M ETHOD
We first explain how to construct the sets of task-oriented and supportive actions and then describe two decision-making strategies used by the robot to perform the task.
A. Action Sets
We define two action sets for the robot to use: task-oriented, A T O ( s ) and supportive , A S ( s ) . A T O includesactions that each move a robot block to its destination, A T O ( s ) = { a = ( b R , d ( b R )) | ∀ b R , l ( b R ) (cid:54) = d ( b R ) } . (1) A S includes the supportive actions. We define a supportive action, a = ( b H , d ) , for a human block, b H , that is closerto the robot. Then we set the closest empty cell to it, whichis also nearer to the human, as its destination d . This way,we balance the cost of the additional action with reducingthe potential for interference while favoring the human’sreference of retrieving objects near them. For e . g ., in Fig.2, A T O = { a , a } and A S = { a , a } . B. Task-Oriented Robot
The task-oriented baseline randomly samples an actionfrom the task-oriented set at a given state, a R ∼ A T OR ( s ) .The goal is to complete the task with the fewest actions.It chooses randomly because all task-oriented actions arenecessary for reaching the goal state. C. Supportive Robot
The
Supportive robot chooses actions using a policy, π containing actions from task-oriented and supportive sets.This policy is an ordered set of actions ranked by theirpriority and is defined by the user before starting the task.Here, we describe the heuristical approach we took to create π for the task with the goal to reflect the utility of supportive actions.We initialize π as an empty list and populate it by iteratingover the following rules until no new action is generated. Wealso initialize B to a list of all the blocks in the grid.1) Return empty if B is empty.2) If a block b iR ∈ B exists such that b iR has no humanblock that might cause a conflict when reaching for it,then pop b iR and return a task-oriented action for it.3) Else, find a supportive action from B that has conflictwith the most robot objects in B .This approach is designed to produce actions that reduce theprobability of collision between the human and the robotwhile trying the minimize the task completion time. It isapplicable to any block configuration.Given a predefined policy, π , the robot checks the listin order and executes the first action that is feasible in thecurrent state s . If no feasible action is found, it defaultsto sampling available task-oriented actions until the goalis reached. We assign a fixed list to π to ensure that theparticipants observe similar behavior from the robot everytrial when studying the effect of supportive actions.Fig.2 depicts an example task with four blocks, task-oriented actions, A T O = { a , a } , and supportive actions, A S = { a , a } . The policy, π , for this scenario is π =( a , a , a ) . Here, a task-oriented action, a , is included firstbecause of the lack of potential goal conflict of block ; thenthe robot takes a supportive action, a , to reduce the potentialinterference of block ; and finally, it completes the taskwith the last task-oriented action a . The planner ignores supportive action, a because block causes no potentialinterference with the robot’s blocks.V. S IMULATED E XPERIMENT
We simulate a scenario with two -link robot arms per-forming pick-and-place actions in D (Fig. 3). Our goal isto observe the effect of supportive actions in an idealizedsetting, without the variance introduced by the participants,or errors in sensing and actuation.We develop an OpenRAVE [15] environment with blocksof two colors scattered on a table. We assign each arm six
Robot (Simulated)
Human
Fig. 3. The simulated 2D environment with two arms, one is a simulationof the human and the other is controlled by the robot policy.TABLE IS
IMULATION R ESULTS
Scenario Robot Mode Task Time (s) Safety Stops
Easy Task-Oriented . ± . . ± . Supportive . ± . . ± . Hard Task-Oriented . ± . . ± . Supportive . ± . . ± . blocks of the same color. The goal for each arm was to bringblocks of their assigned color to the destination area near thearm, highlighted in Fig. 3. We define a × grid on thetable and place the blocks into these cells according to twoconfigurations, easy and hard , as shown in Fig. 4. We con-sider one arm as the robot and the other one as a simulatedhuman. The simulated human chooses task-oriented actionswhile prioritizing closer blocks. We experiment with therobot following both task-oriented and supportive algorithmsfrom Sec. IV. The RRT* [16] implementation in OMPL [17]is used to plan joint-space trajectories. Results.
The two scenarios and two robot modes makefour experimental conditions. We run each of them timesand present the averaged results in Tab. I. The time taken tocomplete the task by the slowest agent is termed Task Time.We also record the number of times the simulated robotwas stopped during the interaction to prevent a collisionand term it Safety Stops. The robot stops and waits for thesimulated human to move a threshold distance away whenthis happens while the human is free to move. We find taskcompletion time to be higher for the supportive robot butthe safety stops are lower in Tab. I. A larger effect due to supportive actions is observed for both metrics in the hard scenario. The supportive robot is always slower than thehuman and although the additional actions cause a longertask time they also reduced goal conflict leading to less than safety stops in the hard scenario. a) Easy (b) Hard Fig. 4. Layout of the easy (left) and hard (right) block configurations, viewed so the human is seated below row A. The human places yellow blocks onthe numbers below row A, whereas the robot is across the table and placing red blocks in Row G. The difficulty is due to the conflict caused by the robotand human reaching for the same space. This conflict exists more in (b) since most of the yellow blocks are in front of the robot’s.
VI. H
YPOTHESES
Following simulation results, we anticipate the robot’sbehavior and the initial block configuration to affect collab-orative performance. We formulate the following hypothesesto test on a user study with a physical robot. H1 Supportive actions will reduce the interference betweenthe agents.
In particular, we expect the supportive ac-tions to reduce the safety stops occurring in the inter-action, especially for difficult scenarios. H2 Supportive actions will reduce the human’s time to com-plete the task.
We expect people to complete the taskfaster when interacting with the supportive robot leadingto more idle time, especially for difficult scenarios. H3 Supportive actions will have a positive effect on thesubjective measures of task performance.
We expectthat participants will prefer the supportive robot as acoworker, especially for difficult scenarios. H4 Changing the initial block configuration would affectboth the subjective and objective measures.
In particu-lar, we expect that participants will find the task moredifficult to perform if the initial block configurationincludes more goal conflicts. We also expect the effectof supportive actions to be more prevalent in difficultscenarios, in general.VII. U
SER S TUDY D ESIGN
We conduct a user study to test the effect of the support-ive actions. The study was approved by Monash University’sEthics Review Board.
A. Independent Variables
We manipulate two independent variables. • Robot mode : { Task-Oriented , Supportive } robots asdescribed in Section IV. • Scenario : { Easy , Hard } block configurations. (Fig. 4)The block configuration in the easy and hard scenariosare designed to cause different levels of goal conflict. Whileboth of them include six blocks, the robot’s blocks in the hard scenario were arranged to be directly in front of thehuman’s. We expect this would increase task difficulty bycausing more interference since both agents need to reachinto the same space. B. Participant Allocation
We recruited subjects aged − ( M = 22 . , SD = 3 . , male, female) for a within-subject study.To reduce order effects, we counterbalanced the order of therobot mode. We kept the scenario order the same, where hard always followed easy . The participants were not informedabout the kind of robot they would be interacting with orhow many types there were. C. Procedure
The experiment took place in a university lab underexperimenter supervision. We seated participants in front ofthe robot as depicted in Fig.5. After reading the explanatorystatement and signing a consent form, the experimenterexplained the task by reading from a script.The participants were assigned yellow blocks and theirgoal was to move these blocks to their destinations accuratelywhile minimizing task time. The start of a turn was signaledon the scanning display in Fig. 5 and both agents performedreaching actions simultaneously, continuing until all theirblocks were in their respective destinations. This concludedone trial and each participant performed four. Participantswere also given three types of surveys, a demographic oneat the start of the experiment, one after every trial, and oneat the end to record their overall experience. A completeexperiment took between and minutes. D. Dependent Variables
We record both objective and subjective metrics.
Objective measures.
We study the effect of supportive actions on task completion time for each agent, the totalnumber of safety stops, as well as human’s idle time ratio.The task completion time is the time an agent takes tocomplete a trial and is easily measured for the robot since weprogrammatically record the time when the robot starts andfinishes an action. For the human, we manually annotate thisusing a video recording of the experiments. We also annotatethe time the human waits for the robot after completing anaction and compute the ratio of the accumulated wait timeover a trial to their total execution time as the human-idleratio. We also count the times the robot has to stop due toproximity to the human as safety stops.
ABLE IIL
IKERT - SCALE COMPOSED OF INDIVIDUAL SURVEY ITEMS WITH C RONBACH ’ S α . (R) INDICATES A REVERSE SCALE . Robot coworker proficiency ( α = 0 . ) I believe the robot accurately perceived my goals.The robot was helpful and/or cooperative.The robot seemed to select the correct object to pick upmost of the time.The robot disrupted me in efficiently performing the task. (R)
I felt uncomfortable with the robot. (R)
TABLE IIII
NDIVIDUAL SCALE ITEMS FROM SURVEY . Individual MeasuresI1
How successful were you in achieving your task? I2 How hard did you have to work to accomplish yourlevel of performance? (R)I3
How much attention did you pay to the robot andits performance during the task? I4 I felt unsafe with the robot. (R)I5
How would you grade the robot as a coworker, overall?
Subjective measures.
Participants answered ten -pointquestions after each trial. Five of these are collected in aLikert-scale that measures robot proficiency as a coworkerand includes statements about the robot’s helpfulness, action-selection, intention-prediction, disruption, etc . . The restof the questions are treated as individual differential scaleitems. We adapt this survey from collaborative HRI studieslike [18]. The Likert-scale (Cronbach’s α = 0 . ) is listedin Tab. II and the individual items are listed in Tab. III.VIII. I MPLEMENTATION D ETAILS
Our user study setup is depicted in Fig. 5 and includes therobot and the human around a table with a checkerboard gridon which we place the blocks. We mount an RGB-D sensoroverhead to detect the blocks and the person’s arm. Thesedetections guide the robot’s action-selection and trajectoryplanning, which are implemented on the Universal Robot 5(UR5) using the Robot Operating System (ROS) [19]. Wealso include a scanning area that instructs the participantabout the destinations for their blocks. Our experiment isfully-autonomous and does not require human intervention.
A. Sensing
The location of the grid is calibrated in the camera frameahead of time using OpenCV [20] and we apply a simplecolor blob detection technique to the RGB image in real-time to localize the blocks.We instructed participants to wear a colored glove cov-ering their arm to allow for its easy detection. We ensuresafety by stopping the robot arm if the user’s hand comeswithin a fixed distance threshold.
B. Robot Control
We implement both task-oriented and supportive robotpolicies for action-selection. For a given goal grid location,we generate waypoints for the robot end-effector to it at
RGBD Camera
UR 5 Manipulator Robotiq GripperScanning DisplayScanning Camera
Fig. 5. The experimental setup a fixed vertical offset from the grid and use the MoveItframework [21] to generate a Cartesian path. This path isfollowed by the robot controller after which it attempts avertical move down to either grab or drop the block and thenmoves back up. Robot joint speed is limited to ( . rad/s )to ensure user safety and comfort.We also included a camera station where participantsscanned blocks and were informed of their destinations aftera short delay. We use this delay to account for the human’shigher relative speed to synchronize human-robot actions.IX. USER STUDY R ESULTS
We compare the independent variables through the objec-tive task performance metrics first and then by participantresponses to the survey. We had to remove the data for twoparticipants, one due to a robot failure, and the other becausethe participant did not follow experimental directions. Thus,in total we analyze ( N = 16) × trials. A. Objective Measures
We analyze some of the objective metrics in Fig. 6.
Safety Stops.
We count the times when the robot has tostop due to its proximity to the human’s arm. We comparerobot types through a Wilcoxon signed-rank test on eachscenario because the data was not normally distributed. Wefind a significant effect due to the supportive robot in the hard scenario ( w = 79 . , p < . ). Fig. 6a shows that the supportive robot had fewer stops in hard affirming H2 . Robot Task Time.
We use a repeated-measure two-way ANOVA to compare the robot’s task completion time.We find a significant effect due to the supportive robot( F (3 ,
60) = 74 . , p < . ) and no interaction. Table IVshows that the addition of supportive actions led to a longerrobot task time. Human Task Time.
We use a Wilcoxon signed-rank testto compare the human’s task completion time due to the a) Safety Stops (b) Human Task Time (c) Human Idle TimeFig. 6. Objective Measures. Box-and-whisker plots of the (a) number of safety stops; (b) time taken by the human to complete the task; and (c) theproportion of idle time spent by the human. Note, T-O refers to the task-oriented robot.TABLE IVT
ASK COMPLETION TIME OF THE ROBOT .Robot Robot Task Time (s)Baseline . ± . Supportive . ± . non-normality of this data. We find no significant effect dueto supportive actions for either scenario. Fig. 6b shows thehuman interacting with the supportive robot is faster but withhigh variance, partly denying H3 . Human-Idle Time.
We use a repeated-measures two-wayANOVA to analyze the human’s idle time ratio as a measureof task fluency. We compute this ratio by accumulating thetime the human waited for the robot to complete an actionbefore they could start the next one and dividing it by thehuman’s task time. We find significant effects due to bothrobot ( F (3 ,
60) = 7 . , p < . ) and scenario ( F (3 ,
60) =5 . , p < . ) types. Fig. 6c shows that supportive robotand hard scenario each led to higher idle time partiallyaffirming H3 . This measure was adapted from [18] whereit was found to be correlated with higher human preference. Supportive actions.
The robot took on average fewer supportive actions in the easy ( . ) scenario than the the hard ( . ) due to fewer goal conflicts. The participants tookonly supportive actions overall and all of them took placein the supportive robot condition. Summary.
The supportive robot confirms H1 in the hard scenario by reducing interference; it partly confirms H3 since human’s idle time is increased, however, the human’stask completion time is not significantly reduced. Also, the supportive robot takes longer to complete this task. B. Subjective Measures
We analyze some of the survey responses in Fig. 7.
Robot coworker proficiency.
We perform a two-wayrepeated-measure ANOVA on the Likert-scale from Table IIand find significant interaction ( F (3 ,
60) = 13 . , p < . ).The normalized responses in Fig. 7a show that participantsprefer the supportive robot in the hard scenario but haveno preference in the easy one affirming H1 for it. They alsoshow that people prefer supportive robot more when the taskdifficulty increases but preference for the task-oriented robotremains similar regardless of task difficulty. Scenario Effect . We use a Wilcoxon signed-rank test tocompare individual scale responses from Table III. We findsignificant scenario effect for both I2 ( w = 0 . , p < . )and I3 ( w = 34 . , p < . ). Fig. 7b indicates thatparticipants find the hard scenario more difficult to perform,affirming H4 . It also leads to the observation that people aremore observant of the robot’s actions in the hard scenario. Safety Perception . We used a Wilcoxon signed-rank testto compare the I4 scale item and do not find any significanteffect due to supportive actions. Fig. 7c shows that partici-pants felt very safe for both robot types in our experiment. Summary . We find that participants prefer the support-ive robot as their coworker in the hard scenario affirming H3 ; also, participants find the hard scenario more difficultand pay more attention to the robot in it, supporting H4 .X. D ISCUSSION
One might think that moving the human’s blocks close tothem would cause people to perceive the robot as helpful andinflate supportive robot’s proficiency. However, our results,which show that the supportive robot is only preferred in the hard scenario, provide evidence for the human’s preferencerelying on the suitability of the robot’s action-selection tothe task.Our results show that supportive actions do not reducesafety stops in the easy scenario. Safety stops are overlapsin agent trajectories and can be caused by an unavoidableconflict between agent goals, uncertainty about each other’sgoal, sensor error, etc. We label a configuration as hard dueto the presence of more goal conflicts; this label does notallude to other sources of overlap. Supportive actions in ourwork were designed to reduce goal conflicts, they lead tofewer stops on the hard task, but will need to be adapted forother sources of conflict to be effective in other scenarios.Hoffman [18] found that collaborative fluency does nottrack task-efficiency in team tasks. Ours is not a team task,however, our results also show coworker acceptance to beseparate from either agent’s task-efficiency. We find sup-portive actions to increase coworker acceptance but reducerobot efficiency. They present a trade-off that needs to beconsidered for designing robot behaviors. For e . g ., if a robotis introduced into a manual process to reduce repetitive a) Likert Scale (b) Individual Measures (c) Safety PerceptionFig. 7. Subjective Measures. Box-and-whisker plots of the (a) normalized survey response to Likert-scale items for the different robot type separatedby scenario; (b) response to measures of subjective task difficulty and attention to the robot for the two scenarios; and (c) safety perception for the robottypes. Note that the leftmost box in (b) and rightmost box in (c) have no height and so appear as a line at . . tasks for humans and increase their job satisfaction, thenits acceptance might play a more important role than itsefficiency. Our methodology helps highlight this trade-off bycombining the subjective and objective impact of supportiverobot behaviors and is applicable to other shared-workspacehuman-robot environments. We consider this methodologyas one of the contributions of our work.XI. C ONCLUSION AND F UTURE W ORK
We introduce interaction-supporting actions and designrobot behavior that selects between these and task-orientedactions by considering the human’s and its own goals. Weimplement it on an autonomous robot and evaluate it in ashared-workspace user study. The results show that this robotincreases human coworker preference in a scenario with moregoal conflicts but decreases efficiency as compared to a robotthat only takes task-oriented actions.Our study illustrates taking actions to support interactionwhile trading off on efficiency in an assembly task. Although,the rationale from Sec. IV can help guide adaptation to newdomains, however, the actions are applicable only to similarscenarios. In future work, we plan to develop a frameworkfor supportive behavior that can perform this reasoning basedon task-specific cost functions.Participants took very few supportive actions towards therobot. We believe their unfamiliarity with the task causeduncertainty about allowed actions. An interesting extensionwould be to apply this to an actual manufacturing task withsubjects who are familiar with it to test the generalizabilityof our findings. We can also improve task naturalness byincreasing robot speed by employing better sensors andmodels for human motion prediction.R
EFERENCES[1] International Federation of Robotics (IFR), “Ifr pressrelease,” 2019, https://ifr.org/ifr-press-releases/news/robot-investment-reaches-record-16.5-billion-usd, Last accessedon 2019-10-01.[2] V. V. Unhelkar, H. C. Siu, and J. A. Shah, “Comparative performanceof human and mobile robotic assistants in collaborative fetch-and-deliver tasks,” in
ACM/IEEE International Conference on Human-Robot Interaction (HRI) , 2014.[3] K. S. Welfare, M. R. Hallowell, J. A. Shah, and L. D. Riek, “Con-sider the human work experience when integrating robotics in theworkplace,” in , 2019. [4] R. J. Stout, J. A. Cannon-Bowers, E. Salas, and D. M. Milanovich,“Planning, shared mental models, and coordinated performance: Anempirical link is established,”
Human Factors , vol. 41, no. 1, pp. 61–71, 1999.[5] G. Hoffman, “Ensemble: fluency and embodiment for robots actingwith humans,” Ph.D. dissertation, Massachusetts Institute of Technol-ogy, 2007.[6] E. Fehr and K. M. Schmidt, “Theories of fairness and reciprocity-evidence and economic applications,” 2001.[7] V. Gabler, T. Stahl, G. Huber, O. Oguz, and D. Wollherr, “A game-theoretic approach for adaptive action selection in close proximityhuman-robot-collaboration,” in , 2017.[8] J. Mainprice, R. Hayne, and D. Berenson, “Goal set inverse optimalcontrol and iterative replanning for predicting human reaching motionsin shared workspaces,”
IEEE Transactions on Robotics , vol. 32, no. 4,pp. 897–908, 2016.[9] S. Li and J. A. Shah, “Safe and efficient high dimensional motionplanning in space-time with time parameterized prediction,” in , 2019.[10] K. P. Hawkins, S. Bansal, N. N. Vo, and A. F. Bobick, “Anticipatinghuman actions for collaboration in the presence of task and sensoruncertainty,” in , 2014.[11] S. Nikolaidis, R. Ramakrishnan, K. Gu, and J. Shah, “Efficient modellearning from joint-action demonstrations for human-robot collabora-tive tasks,” in
ACM/IEEE international conference on human-robotinteraction , 2015.[12] A. Cherubini, R. Passama, A. Crosnier, A. Lasnier, and P. Fraisse,“Collaborative manufacturing with physical human–robot interaction,”
Robotics and Computer-Integrated Manufacturing , vol. 40, pp. 1–13,2016.[13] H. S. Koppula and A. Saxena, “Anticipating human activities usingobject affordances for reactive robotic response,”
IEEE transactionson pattern analysis and machine intelligence , vol. 38, no. 1, pp. 14–29, 2015.[14] M. C. Gombolay, R. A. Gutierrez, S. G. Clarke, G. F. Sturla,and J. A. Shah, “Decision-making authority, team efficiency andhuman worker satisfaction in mixed human–robot teams,”
AutonomousRobots , vol. 39, no. 3, pp. 293–312, 2015.[15] R. Diankov, “Automated construction of robotic manipulation pro-grams,” Ph.D. dissertation, Carnegie Mellon University, RoboticsInstitute, August 2010.[16] S. M. LaValle, “Rapidly-exploring random trees: A new tool for pathplanning,” 1998.[17] I. A. S¸ucan, M. Moll, and L. E. Kavraki, “The Open Motion PlanningLibrary,”
IEEE Robotics & Automation Magazine , 2012.[18] G. Hoffman, “Evaluating fluency in human–robot collaboration,”
IEEETransactions on Human-Machine Systems , vol. 49, no. 3, pp. 209–218,2019.[19] M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs,R. Wheeler, and A. Ng, “Ros: an open-source robot operating system,”2009.[20] G. Bradski and A. Kaehler,