Optometrist’s Algorithm for Personalizing Robot-Human Handovers

The handover motion was developed using the Sawyer Robot’s motion controller. The reach actions were based on ’ROS MoveIt!’. The robot end-effector was commanded to reach the necessary coordinates in Cartesian space. A handwritten script controlled the robot’s actions and phases of the handover.

The robot picked up the object from a table (pick-up location) and moved to the starting location (both poses predefined and fixed throughout all experiments). From there, it began the ’reach’ phase of the handover and moved towards the user (location of the handover - proximity, side, and height $[x,y,z]$ - as well as the speed of the reach $V_{max}$ was tunable through the OA). Once the robot stopped, it waited for the user to initiate the object transfer. The user could take the object from the robot’s grasp by overcoming a force threshold ($F_{min}$, tunable through OA). Once the object was released, the robot moved back above the pick-up location. Meanwhile, the user had to restore the object to the pick-up location and prepare for the subsequent handover. We chose this method of determining a handover location according to a fixed pose defined by parameters over others (like computer vision), as it standardizes the handover and its tuning for each sample and each participant. Also, the need to tune a pose in the robot’s reference frame provides us with a non-intuitive parameter.

The parameters that we tuned were as given in Table I. For each parameter, a lower and higher limit and a step size were determined empirically before the experiment, resulting in five sets of roughly ten parameter values each - to be tuned by the OA.

The Optometrist’s algorithm (OA, see Algorithm III-B) was developed and used to determine a personalized set of parameters using a technique similar to an optometrist determining the correct power of one’s eye-glasses. The user was shown two options, variations of a single parameter, and was asked to choose the preferred option. Once the tuning of a parameter converged, the OA moved on to the next one. One of the options was the handover preferred in the previous step, and the other was a new option (with the obvious exception of the very first step). After the user chose an option, the algorithm stepped, varying the parameter by one step size ($s$). The previous choice decided the direction of the step. If the new option was chosen, the next step was taken in the same direction as the new option. If the same option was chosen again, then the direction for the next step was flipped. This allowed us to parse and present the entire parameter range to the user, possibly accessing parts of previously disregarded parameter space if needed. The process finished when the difference between the two options to be presented to the participant became smaller than the step size or when the same option was chosen four times in a row (this second stopping criterion is omitted in Algorithm III-B for readability). We predefined the order in which we tuned the parameters as Speed, Position $(x,y,z)$, and Force. In a few iterations of this exercise, the algorithm converged to the set of parameters that the user preferred the most.

The Optometrist’s Algorithm {algorithmic} \State$L\leftarrow Min.$ \State$H\leftarrow Max.$ \State$M=avg(L,H)$ \State$s\leftarrow Step$ \State$handover(L);$ \CommentRun handover with parameter L \State$handover(H);$ \State$choice\leftarrow user\_chose(L,H)$ \If$choice=L$ \State$option1\leftarrow L$ \State$option2\leftarrow M$ \Else\State$option1\leftarrow H$ \State$option2\leftarrow M$ \EndIf\While$|option1-option2|>s$ \State$handover(option1);$ \State$handover(option2);$ \State$choice\leftarrow user\_choose(option1,option2);$ \If$choice$ is $option1$ \If$option1>option2$ \State$H\leftarrow H-s$ \State$option2\leftarrow H$ \Else\State$L\leftarrow L+s$ \State$option2\leftarrow L$ \EndIf\State$option1\leftarrow option1$ \ElsIf$choice$ is $option2$ \State$temp\leftarrow option2$ \If$option1>option2$ \State$option2\leftarrow option1-s$ \Else\State$option2\leftarrow option1+s$ \EndIf\State$option1\leftarrow temp$ \EndIf\EndWhile

The OA presented here is optimized for small-scale problems with fewer parameters and a limited range. However, it is possible to tune several thousand parameters at a time, even with more extensive ranges, by introducing probabilistic steps and tuning sets of dependent parameters [13]. In any case, setting up the parameters and their ranges takes some expert knowledge and effort.

The participants performed robot-human handovers during two experiment phases, preceded by short practice sessions each. A tuning and an evaluation phase were performed sequentially in a single sitting of about 20 minutes. During the experiment, the participants were seated in a fixed chair in front of the robot, with a table being the shared workspace of the human-robot team (Fig. 1). To obtain qualitative data, the participants were asked to complete questionnaires before and after the experiment (see Sec. IV-C). The first practice session of five handovers before the tuning phase was for participants to get used to the robot’s actions and the task. The parameters used for these handovers were a fixed set of near-average parameters, the same for each participant.

In the tuning phase, the participants interacting with the OA were told what aspect of the handover they were tuning in layperson’s terms (speed, position, and the force of the grip). The participants had to verbally announce their choice to the experimenter, who then used a simple interface to register the choice. In the second practice session of five handovers (between the tuning and the evaluation phases), the participants were asked to perform the personalized handovers. The handovers performed in these practice sessions were used to study the differences between the handovers before and after tuning. The participants were again shown two options in the evaluation phase. One of the options was their personalized handover, and the other was a different one - similar to the preferred handover, but with slight noise added to one of the parameters. Participants were asked to identify their tuned handover. Note that as only one parameter was varied at a time and within a limited amount (1-2 times the step $s$), the handovers were still relatively similar to each other, and it was more challenging for the participants to find their own handover (compared to varying more parameters and in a broader range).

The experiment was performed with $N=30$ participants (Ages: $19-38$, $(M)28.8\pm(SD)4.6$ years, $50\%$ Females), all of whom volunteered to be in the experiment for no compensation. All the participants were healthy university students or researchers. The participants were from a wide range of professional and educational backgrounds. The experiment was approved by the ethical committee of the Ben-Gurion University’s Department of Industrial Engineering and Management.

The participants were asked to complete the ”Technology Adoption Propensity (TAP)” [14] and ”Negative Attitude towards Robots Scale (NARS)” [15] questionnaires before the experiment. After the experiment, the Technology Acceptance Model (TAM)[16] was used to study the acceptance of the personalization task based on factors such as ’Ease of Use,’ ’Perceived Usefulness,’ ’Attitude’ of the users, and their ’Intention of Using’ this method in the future. The questionnaire used was based on questionnaires established in earlier studies[17],[18]. The ’Effort’ component of the ’Ease of Use’ variable was evaluated using the NASA-TLX Questionnaire [19] for perceived workload and effort. For ’Trust’, the ’Competency’ and the ’Reliability’ sub-scales of the Multi-Dimensional Measure of Trust (MDMT)’ [20] were used. All evaluations were based on a 7-point Likert scale (unless stated otherwise).

We also studied the subjective perception of factors such as safety and fluency of the handovers. Safety was evaluated using the Godspeed Questionnaire for Perceived Safety [21], and the fluency questionnaire was based on the works of Hoffmann [22] and Paliga et al.[23]. The subjective questionnaire is detailed in Fig.3

The experiment had a very high handover success rate ranging from $94.83\%$ to $100\%$ ($mean=99.46\%$), including the handovers during tuning.

Of the failures, three occurred because the Robot’s motion controller failed to generate a collision-free trajectory. Four failures were due to false triggering of the transfer. Once, a participant dropped the object with no fault of the robot. For the failures during the tuning phase, the pair of options was repeated to continue tuning.

The distribution of parameters preferred by the users is shown in Fig.4. It is clear from the data that different participants enjoy different parameter values and can tune for them. The participants preferred to tune the handovers to be at a proximal handover location rather than the mean parameter value ($p<0.01$). But most people decided not to let the robot come too close. While choosing the side of the handover, $73.33\%$ participants preferred to keep the handover location at the center. Of the people that preferred to choose either the left or right side ($8$ out of $30$), all decided to choose the side of their dominant hand. Most participants choose to keep the handover height lower than the mean (starting) value of the parameter ($p<.001$). The users preferred handovers faster than the mean speed ($p<0.001$). While this parameter can also be task-dependent, even non-experienced users choose to go for higher speeds in robots. Participants preferred low values of threshold trigger force ($p<0.02$) compared to the parameter range’s mean. The users tried to minimize their efforts while performing the handovers, as reported in [22]. Many users admitted that they would have liked to choose a lower force than the minimum of the parameter range. However, in this study, we were limited by the design of the controller (robustness).

The entire tuning process took between $10-12$ minutes and $18$ and $21$ steps ($mean=19.2\pm 1.5$), amounting to twice the number of handovers performed. Tuning the handover height required the most steps ($4.1\pm 0.9$ and $143\pm 32$ sec) while tuning the side of the handover location (left or right) required the least steps ($3.2\pm 0.5$ steps and $110\pm 21$ sec). Despite every user needing to perform approximately $40$ handovers for the personalization, the users perceived this activity as a low workload on the NASA-TLX Questionnaire (score$=2.68\pm 1.49$ on a 7-point Likert scale).

Most participants could identify their handover between variations in speed ($mean\ score=0.75,$ out of $1$), followed by variations in threshold force values required to trigger the transfer ($mean\ score=0.58$ out of $1$). For variations in handover location, the participants performed the worst ($mean\ score=0.53$ out of $1$). In five attempts, the average participant scored $3.19\pm 0.92$. The participants who performed the best were those who had tuned the values to the extreme ends of the parameter range.

The users evaluate the personalization process very positively through the subjective questionnaire (Fig.5). The process is primarily perceived as useful ($5.44\pm 1.29$) with $19$ out of $30$ users. $21$ out of $30$ users agreed that the method was easy to understand. The users also showed a positive attitude towards this method, especially in its ’Engagement’ component, with $15$ out of $30$ participants strongly agreeing to have been completely focused on the activity. Users ($87\%$, $26$ out of $30$) generally agreed ($5.16\pm 1.69$) that they might use this method in the future. Of the four users who disagreed, one reported that the tuning process did not produce a desirable handover. Two users mentioned that the process was ”too repetitive” and was ”not enjoyable”, while one user felt that ”the robot did not respond to their feedback”.

On the Multi-Dimensional Measure of Trust (MDMT) scale, the user scored the robot highly on both the Reliability ($5.66\pm 1.41$) and the Competency sub-scales ($5.49\pm 1.29$). The users also perceived high robot safety, scoring $4.14\pm 0.86$ on a five-point scale of the Godspeed questionnaire. The user perception of trust and safety did not change even for those participants who experienced one of the few handover failures. Few users reported that they were at times surprised by some of the options that were ”too fast” or when the robot ”came too close”. However, all the users rated the task highly on the trust and safety scales. A low correlation was observed between success rate and safety ($r=0.20$) and between success rate and trust ($r=0.15$).

Our metrics have suggested a high perception of both subjective fluency ($5.28\pm 1.22$) and improvement in fluency ($5.59\pm 1.10$) after tuning. The significance of all the metrics was evaluated using a two-tailed test. A significant increase in H-IDLE ($p<0.01$) and a highly significant decrease in F-DEL $(p<0.01)$ was observed after personalization. Both these observations correspond to an increased objective fluency after personalization [22]. Comparing fluency metrics to subjectively perceived fluency, there was no correlation between either change in H-IDLE ($r=0.096$) or F-DEL ($r=-0.092$). This implies that the handover fluency improved objectively, but not all participants could perceive it. An inconsistent subjective perception of fluency in a proactive coordination task, such as ours (the robot did not wait for a signal from the human before starting the handover), is already reported by Huang et al.[25].

There was no significant change in C-ACT ($z=1.17,p>0.1$), which was expected, as the delays programmed in the controller were constant, to make the tuning interaction more consistent. After personalization, R-IDLE showed a highly significant decrease ($p<0.01$). A major factor contributing to R-IDLE was the robot waiting for the user to initiate the object transfer. Initially, as the handover location was not personalized and the force to trigger the transfer was not preferred by most users, this phase took longer. This metric can also be interpreted as task efficiency in terms of robot usage during the task. This task efficiency improves significantly after personalization.