The Impact of Route Descriptions on Human Expectations for Robot Navigation

It has long been suggested that context-aware systems should be able to explain to users “what they know, how they know it, and what they are doing about it” [4]. Research into context-aware predictive systems has found that people use explanations to clarify their uncertainties about what the system is doing or why [31, 32]. Many different algorithms have been proposed to provide that information to system users [2], and research has shown that people rely more on decisions of machine learning systems that are explained compared to ones that are not explained [8]. Recently, Amir and Amir [3] proposed a method of automatically summarizing previous behavior by agents that was found to help people more appropriately assess the agents’ skills.

More recently, it has been suggested that explanations could be one way to help people understand what to expect from autonomous robots as they act in human environments [45]. Many people have proposed different types of explanations for different types of robot behavior. For example, Hayes and Shah [23] proposed explaining why, why not, and how explanations of robot policies. Chakraborti et al. [11] proposed several types of explanations that provide the minimal information needed to “patch” an observer’s mental model or that are executable. Other work has proposed explaining the cost functions related to Markov Decision Processes to help people understand the trade-offs that their robots are making [43]. Specifically related to robot navigation, verbalization has been proposed to summarize robot paths at different levels of detail for different types of users to understand where the robot has been [41]. More recent developments suggest that there are ways to reduce the size of an explanation by finding important features in an abstract policy [46]. Our goal is to understand how different explanations of robot route navigation behavior impact user expectations of that behavior.

Rather than reinvent suitable explanations of robot route navigation, we draw significantly from the human-generated route instruction literature to develop our explanations of robot navigational routes. Much work has investigated route instructions and their ability (or inability) to help readers accurately navigate to desired goal locations. This research often refers to landmarks, used in the colloquial sense of “a prominent or conspicuous object on land that serves as a guide... a distinguishing landscape feature marking a site or location” [1]. Researchers have found that successful route instructions (those that readers can understand most frequently) include “minimal and essential landmark information” and a set of instructions for what to do at those landmarks [13, 15, 16]. What constitutes minimal, essential information is a reflection of the map or context. The specific selection of such key information can be critical to explanation quality, as is an explanation’s ability to contrast a described route from the alternatives [33]. In contrast, GPS navigation technology typically uses road names, distances, and turns. The prospect of maintaining valid databases of landmarks and updating them when changes occur is expensive, and road names change less frequently than landmarks [6]. Comparisons between instructions with and without landmarks have shown that people are significantly better at following instructions with landmarks than without them [14].

The question of what comprises minimal information seems to depend on prior knowledge of the environment. When readers are newer to the environment, they respond best to distance and landmark instructions as they develop an egocentric (first-person view) cognitive map of the environment [34]. However, more knowledgeable readers, who have developed a richer, allocentric (global bird’s-eye view) map with distances and routes between more landmarks, tend to think about instructions hierarchically as routes between well-known landmarks [34]. With similar reasoning, some people choose to provide turns in an allocentric way using cardinal directions (north, south, east, and west) while others give directions in an egocentric way as if the instructions were being executed by the reader (left, right) [12]. Interestingly, while “minimal” instructions of all necessary turns are sufficient, people tend to provide additional information at times when the path is “tricky” in some way or to help validate that the user is or is not traveling in the correct direction (e.g., “If you see the school, you’ve gone too far”) [24].

With some widely-understood patterns (e.g., use of landmarks) and some diversity in instructions (e.g., egocentric vs allocentric instructions, specificity of instructions), we developed several different explanation types to understand their impact on readers’ expectations of the routes they are reading about. Details about the particular route navigation explanations that we study are provided in Section 3.2.

Understanding user expectations of robot behavior is a challenging task. Research often implicitly assumes that observers start with no knowledge of the robot behavior and that explanations should therefore completely describe the robot’s behavior [23, 41, 43]. In contrast, Chakraborti et al. [10] gave people a starting mental model (description of the problem) and then modified that model to test what people chose to explain. In this work, we provided people with varying, potentially incomplete descriptions of robot behavior and then measured their prediction of and surprise at the true executed behavior. This allowed us to test the impact of the explanations on the participants’ expectations directly rather than assuming what the reader interpreted from the explanation.

In particular, we gave participants two tasks that interfered visually [37] and measured on which task participants chose to focus. Work on surprise and attention has shown that people will take their focus off of one task to dwell on a surprising event [6]. Surprising events are those that deviate from an observer’s expectations; in our case, this would be a robot taking a path that is notably different than the one the participant expected. This idea is similar to measuring neglect tolerance—a human monitor’s ability to let a robot act on its own autonomously without watching its behavior—except that the participants in our study were required to monitor the behavior occasionally and the robot acted autonomously whether or not they monitored it [21]. By measuring the decrease in performance on one task in our study, we can understand how surprising the path is given the explanation.

We used an Anki Cozmo robot, which is a small, expressive robot that moves across flat surfaces on two treads. It has a pixelated display to convey eye movements and expressions. Its software development kit (SDK) allows for programming a large range of behaviors, including navigation.

To give Cozmo a navigation context that participants would intuitively understand, we used a city map carpet (Figure 1). The carpet measured 4 ft by 6 ft and contained unlabeled streets and 18 labeled landmarks. Prior work indicates that people remember instructions better when grounded on landmarks, so this setup is sufficient to support our goals [13, 15, 16]. The streets contained crosswalks (white stripes). We created six paths in various directions for Cozmo to traverse that crossed five or six crosswalks each (Figure 1). The paths varied in their number of turns and directness. The paths contained crosswalks, because identifying crosswalks and obstacles moving in them are of key importance when a robot navigates real streets. Participants were instructed to monitor the robot as it traveled and indicate when it crossed a crosswalk to simulate training a crosswalk or obstacle detector.

To make navigable paths, we created a graph of streets, intersections, and landmarks and computed straight and turn actions. We used Cozmo’s SDK to pre-program the paths and create action lists computed from the graphs. The paths were open-loop (as the Cozmo was not actively sensing its location), but the robot was able to execute the paths while staying in the roadways if it moved slowly. Though the robot could be viewed in person for future experiments, we video-recorded its paths from above for this online experiment and displayed them at \(2.5\times\) speed to counteract the robot’s slowness.

Many algorithms have been proposed for explaining or describing robot navigation behavior. For this context, we had to consider that our algorithm only has access to two attributes: a map with labeled landmarks to reference but no street names and a path (i.e., a sequence of actions from the start to the goal). Given the information available, we chose to focus our explanations on widely-trusted patterns of human-generated navigation instructions. We created four different route explanation types that we could implement automatically and one human-generated explanation. In total, there were six explanation conditions:

We had two hypotheses about the effects of explanation type on participants’ path predictions, their ability to perform two simultaneous tasks, and subjective measures of trust in the robot. First, we expected that providing any explanation for the robot’s behavior would increase these three measures compared to Baseline (H1). Second, we predicted that shorter and more commonly occurring (i.e., passing by or egocentric turns) explanations would increase the aforementioned three measures more than longer, complex explanations with turns in them (H2).

In the dot-tracking task, participants were asked to use their mouse cursor to track a quickly moving red dot 50px wide (Figure 2(b), left) on a 600 \(\times\) 450px area. The dot was programmed to navigate to goal x and y locations within the bounded area, bouncing off of walls when necessary. The dot position updated one pixel on both axes once every 0.2 s. Each time the dot got to the x or y goal, it computed a new goal on that axis and moved again. As a result, the dot frequently changed directions at random rather than moving predictably, making it harder to track.

The dot started moving when the participant pressed the space bar. As feedback during the task, participants received 1 point every 0.2 s if their cursor was anywhere on the dot and lost 1 point if it was not. The speed and unpredictability of the dot meant that the task required substantial concentration to receive positive points. Participants needed a positive score (\(\gt\)0) during all six paths to be included in the dataset.

In the robot navigation task, participants were told to press the “z” button each time the robot entered a crosswalk. They were not told how many crosswalks to expect. Participants viewed a 600 \(\times\) 680px video of a real Cozmo navigating over the city carpet, recorded from above (Figure 2(b), right). Participants must occasionally look at the robot’s progress throughout the video to perform this task successfully. The robot is small and white, and the multi-colored map makes the robot difficult to find at a brief glance. Thus, participants had to completely shift their visual attention between the dot-tracking task and this task to complete both tasks with reasonable success. They could look at the video less frequently if they could predict the path and anticipate where the crosswalks would occur. The surprise-attention literature says that participants will tend to look at the video more frequently and dwell on it longer if they are surprised by the path or unsure of where the robot will travel next.

When the space bar was pressed, the video started playing and the dot started moving on the adjacent panel. Participants’ key presses were logged. If they pressed “z” during a crosswalk, then they received 1,000 points on their crosswalk score. They could only receive the points once per crosswalk. We removed participants who did not press “z” for at least one crosswalk on every path.

After providing informed consent, participants completed a questionnaire asking for their age, gender, occupation, native language, country where they spent the most time as a child, whether they had any robots or pets and if so what kind(s), and the Ten-Item Personality Inventory [22].

Next, participants were given an overview describing both tasks. The overview specifically included the participants’ goal for the experimental trials: “You will earn points for each correct button press and also based on how long your mouse stays on the red dot. Your goal is to earn the best combined score of tracking the dot and training the robot for each robot path. You will be disqualified for dot-tracking scores lower than 0 or less than 500 points for crosswalks on each page.”

Then, participants were taken to a practice website with more detailed instructions about the task procedure. First, they were told that they would see a city map with labeled landmarks and a description of a robot’s path through that map. The instructions said, “Study the explanation and answer the questions on the next page.” They were given an example explanation next to the map and told upon advancing to the next page, “Click 10 points in order along the path that you think the robot will take from the start to the end.” When they were ready to continue, participants were asked to click the start and end locations of the path and then click to place dots on 10 intermediate points showing their predicted path. If they got the start and/or end landmark(s) wrong, then they were instructed to try again and reread the explanation (and then repeat the clicking). We did not check the predicted paths for accuracy during the experiment. Next, participants were told that they should track the red dot with their mouse and practice pressing the “z” button at any time (interface shown on Figure 2(b)). They pressed the space bar to start and could click “Next” to move to the next page once they had pressed “z” at least once. Participants were reminded not to navigate away from the website during the tasks as the tasks would continue and they would be disqualified for missing data.

After completing the practice trial, the participants started the study. The six paths were shown in a random order. For each path, the participant (1) read the explanation, (2) clicked the start and end landmarks and 10 intermediate points along the predicted path, and (3) pressed space to start the two tasks. (The exact instructions said, “Start the experiment. Track the dot using the mouse and press ‘z’ when the robot crosses crosswalks. Press ‘space’ to begin.”) When the video stopped playing, a “Next” button appeared for continuing to the next path. When all six paths were completed, participants were taken to a final survey that included 13 rating questions and a request to estimate the total number of crosswalks that the robot crossed.

For each path, our website logged participant ID, path ID, path order, every click and key press on the interface (including the start/end and predicted path points), the location of the dot and mouse, and the time of the video every 0.2 s while the dot/robot was moving. At the end of each path (including the practice), we logged the data to a file on our server. From these log files, we computed the measures that we used to analyze differences in trust in experimental conditions:

Path Integral: Additionally, we computed the user’s expectation of the robot’s path using their 10 clicks by connecting the clicks in sequence. We found that about 2% (20 paths out of 180*6 = 1,080) of the clicked paths were noisy, where participants clicked in the order they remembered or inferred parts of the path instead of in order from the beginning to the end. We corrected the paths by sorting the points from top to bottom or vice versa, and we smoothed the sorted paths to ensure that they never went backwards away from the goal. To do this, we first aligned points vertically that were within 150px of each other so that they shared a y value and then computed the longest sequence of points with increasing y values. We then computed the integral as the distance from the smoothed paths to the robot’s path at each pixel height from the top to the bottom of the map. The corrections improved (reduced) path integral scores for all participants.

Subjective Trust: The rating questions at the end of the experiment used 7-point scales from Strongly Disagree to Strongly Agree. Participants rated the following statements: “I am confident in the robot” (confidence); “The robot is dependable” (dependability); “The robot is reliable” (reliability); “The robot’s behavior can be predicted from moment to moment” (predictability); “I can count on the robot to do its job” (count on); “The robot was malfunctioning” (malfunctioning); “I trust this robot” (trust this robot); “I trust robots in general” (trust general); “I will not trust robots as much as I did before” (not trust general); “I could not focus on the tracking task, because the robot needed my attention” (not focus); “I spent more time watching the robot than on the tracking task” (time on robot); “It was hard to complete the tracking task while watching the robot” (hard to track); and “If I did this task again, I would spend more time watching the robot” (future time on robot). Some items assessing trust were adapted from Jian and colleagues [26] (confidence, reliability, predictability, trust this robot) and Muir [35] (dependability, reliability, predictability).

We recruited 242 participants from the Prolific.co recruitment website and paid them 5USD for their time. To participate, site users had to have identified themselves as 18 years of age or older, fluent in English, and having normal or corrected-to-normal vision. Data from 47 participants had to be removed for technical issues that were fixed partway through data collection. In order for the remaining individuals’ data to be included, each participant had to press “z” no more than 50 times throughout the experiment, identify at least one crosswalk correctly per map, have a dot score greater than zero for every map, and not navigate away from the page or lose connectivity for more than 10s or for any period of time that overlapped with a crosswalk on each map. Thirteen participants were eliminated for these issues.

In total, data from 182 participants were included in the analysis. Of these participants, 120 were male, 61 female, and 1 genderfluid; 82 were ages 18–23 years, 42 were 24–29, 43 were 30–39, 10 were 40–49, and 5 were 50–59; and 50 said they spent most of their childhood in North America, 10 in South America, 115 in Europe, and 7 on other continents. There were 30 participants who successfully completed each condition except for Baseline, which had 32. This research was approved by our Institutional Review Board.

We analyzed differences in Path Integral based on the explanation type. Figure 3(a) shows the mean numbers of pixels (and standard deviations) enclosed between the two paths out of a total of 3.5M total pixels comprising the image. (A lower integral is better as there are fewer pixels separating the paths.) The average Path Integral over all of the data was 504,034. Good estimates were under 400,000 pixels, and the worst estimates were over 1M pixels.

Explanation had a significant effect on the Path Integral (F(5,5) = 51.12, \(p\lt 0.001\)). A post-hoc Tukey’s Honestly Significant Difference (HSD) test with an \(\alpha = 0.05\) showed that the Baseline condition (M = 840,361, s.d. = 148,089) was significantly worse than all of the other explanations (M = 431,000, s.d. = 148,081). Additionally, the Passing By explanation (M = 373,926) was significantly better than the Avoids explanation (M = 494,001). No other differences were found.

Path also significantly affected Path Integral (F(5,5) = 47.50, \(p\lt 0.001\)). A post-hoc Tukey HSD test showed that participants were significantly worse at predicting paths 5 (M = 779,238, s.d. = 116,112) and 6 (M = 625,955) compared to the others (M = 398,750). Those two paths were less direct from the start to the goal than the others, and participants who did not understand or remember the explanation were not able to infer it from the start and the goal.

Finally, the interaction between Explanation and Path was significant (F(25,25) = 15.04, \(p \lt 0.001\)). A Tukey HSD test focused on the effects of Explanation for each Path. For Path 3, the Baseline (M = 635,142) was significantly worse than Passing By (M = 331,324) and Avoids (M = 330,513). For Paths 1, 4, and 5, the Baseline was significantly worse than all other Explanations. We found no significant effects of Explanation on Paths 0 and 2.

Based on these findings, we conclude that an explanation does help align user expectations of robot behavior.

Next, we analyzed measures of dual-task performance and interference (Crosswalk Mouse-on-Dot proportion, shown in Figure 3(b)). If participants were attending to the robot task, then their performance on this metric should have dropped. Our REML analysis of Mouse-on-Dot proportion during crosswalks found that Explanation type was statistically significant (F(5,5) = 4.83, \(p \lt 0.01\)) and Path was statistically significant (F(5,5) = 2.54, \(p \lt 0.05\)), while the interaction effect was not statistically significant. A Tukey HSD post-hoc analysis of Explanation type showed that Passing By (M = 83.62% on dot, s.d. = 4.26%) was significantly better than the Baseline (80.36%), Allocentric Turns (79.64%), and Human-Generated explanations (78.63%). No other pairs were statistically significant. During turns and control areas, our REML analyses also found that Explanation type was statistically significant (F(5,5) = 4.62 and F(5,5) = 4.98, respectively, \(p \lt 0.01\)), while Path and the interaction effect were not statistically significant. A Tukey HSD post-hoc analysis of turns showed that Passing By (M = 85.63%, s.d. = 3.12%) was significantly better than Avoids (83.06%), Allocentric Turns (82.88%), and Human-Generated explanations (82.25%). Similarly, the Tukey HSD post-hoc analysis of the control times also showed that Passing By (86.17%, s.d. = 3.23%) was significantly better than Avoids (83.25%), Allocentric Turns (83.34%), and Human-Generated explanations (82.72%). Additionally, Egocentric Turns (85.13%) were significantly better than Human-Generated explanations. Overall, we find that Passing By explanations significantly improve dot tracking compared to Baseline during crosswalks and relative to other explanations at other times. Participants who had that explanation spent more time tracking the dot rather than observing the robot’s behavior.

We tested whether more dot tracking resulted in worse robot tracking. We first analyzed the average time to the first z-press per path. Path significantly affected z-press time (F(t,t) = 44.80, \(p \lt 0.001\)), Explanation had a marginally significant effect (F(5,5) = 2.07, \(p = 0.067\)), and there was no significant effect of the interaction between the two dependent variables. In order from shortest to longest, Path 2 z-press times were significantly lower (M = 2.44 s, s.d. = 0.14 s) than Paths 0 and 1 (M = 2.55 and 2.59, respectively), which in turn were significantly lower than Path 3 and 4 (M = 2.7 and 2.72, respectively), which in turn were significantly lower than Path 5 (M = 2.91). On inspection of the paths (Figure 1), this order reflects the complexity of the paths and their directness from the start to the goal. A test of the z-press proportion did not show any statistical significance from any of our independent variables. Participants were good at identifying crosswalks at the required times—over 95% for each explanation type and each path.

Overall, participants chose to allocate attention to the robot monitoring task at the cost of their performance on dot tracking.

Finally, we evaluated each participant’s final survey to understand how the explanations affected their perceived trust and predictability. We performed a factor analysis on all 13 questions to determine which responses were correlated and how individual items could be grouped into scales. Four factors were identified: Credibility (F1) included confidence, reliability, dependability, predictability, count on, malfunctioning (reverse scored), and trust this robot; Distraction (F2) included not focus and hard to track; General Robot Trust (F3) included trust general and not trust general (reverse-scored); and Future Attention (F4) included future time on robot and time on robot. Of these four factors, only Credibility (F1) showed a significant effect of Explanation condition (Wilcoxon/Kruskal-Wallis test for non-parametric data, \({\chi }^2 = 12.144, p = 0.033\)). Pairwise comparisons were performed using the Steel-Dwass method to correct for multiple comparisons and found that the Baseline condition was significantly worse than Passing By and than Egocentric Turns (all p \(\lt 0.05\)); no other significant differences were found.

Examining the individual questions within F1, Wilcoxon/Kruskal-Wallis tests showed that Explanation condition had a significant main effect on confidence (\({\chi }^2 = 16.069, p = 0.0067\)) and predictability (\({\chi }^2 = 21.556, p = 0.0006\)). For confidence, Steel-Dwass pairwise comparisons showed that Baseline received significantly lower scores than Avoids and than Egocentric Turns. For predictability (Figure 3(c)), Baseline received significantly lower scores than Passing By, Egocentric Turns, and Human-Generated. No other individual questions showed significant main effects of explanation (\(\alpha = 0.05/7\) with Bonferroni corrections).