One of the first questions that comes up when designing a robot tutor is whether the robot should take the role of a teacher or a peer. Research on children’s language acquisition has demonstrated that children learn more effectively from an adult who can use well-defined pedagogical methods for teaching children using clear directions, explanations and positive feedback methods (Matthews et al., 2007). However, designing and framing the robot as an adult tutor has the disadvantage that children will form expectations about the robot’s behavior and proficiency that cannot be met with current technology (Kennedy et al., 2015). Due to technological limitations of the robot and underlying software, communication breakdowns are more likely to occur than with a human. For a peer robot introduced as a fellow language learner, breakdowns in communication are more acceptable. Moreover, interacting with robots acting as peers is conceived as more fun (Kanda et al., 2004), allows for learning-by-teaching (Tanaka and Matsuzoe, 2012) and has a proven to be efficient in teaching children how to write (Hood et al., 2015). Furthermore, there is some evidence that children’s learning can benefit from interacting with peers (Mashburn et al., 2009). Given these considerations, we believe it is desirable to frame or introduce the robot as a peer and friend, yet design its interactions insofar possible based on pedagogically well-established strategies to scaffold language learning.

To implement effective tutoring, the robot needs to interact with children in multiple sessions, so they have to be motivated to engage in long-term interactions with the robot. Establishing common ground between child and robot can contribute to this (Kanda et al., 2004), but first impressions to establish trust and rapport are also crucial (Hancock et al., 2011).

Despite the wealth of studies regarding the introduction of entertainment robots as toys to children (e.g., Lund, 2003), surprisingly little research has been conducted on designing protocols on how to introduce a robot tutor to a group of preschool children. Fridin (2014) presents one exception, and found that introducing a robot tutor to children in group sessions improved subsequent interactions compared to introducing the robot to children in individual sessions. Another study by Westlund et al. (2016) found that the way a robot is framed, either as a machine or a social entity, affected the way children later engaged with the robot. They concluded that introducing the robot as a machine could create a more distant relation between child and robot, thus reducing acceptance. We therefore decided to frame the robot in our project as a social playmate for the children and introduced the robot in a group session. However, the NAO robot is slightly taller and more rigid than the fluffy huggable Tega robot, which Westlund et al. (2016) used, and we observed that some 3-year-old children were somewhat intimidated by the NAO robot on their first encounter. Such a first impression of the robot could reduce the trust that the child had for the robot, which could negatively affect their willingness to interact with the robot in the short-term, but also in the long-term. To develop a successful first encounter and to build trust between the child and robot, we designed the following strategy for introducing the robot to 3-year-old children at their preschool.

Pilot studies revealed that some children got anxious when the robot was introduced and then suddenly started to move. To familiarize children prior to their first encounter with the robot, it is therefore advisable to prepare them well. For our study, we sent coloring pages of the robot to the preschools during recruitment and asked the pedagogical assistants to talk a little bit about the robots to the children. About 1 week before the experimental trials, the experimenters introduced the robot in class during their daily “circle time”, as this provided a safe and familiar environment with the whole group in which the pedagogical assistants usually introduce new topics or new activities. One experimenter first introduced the robot by telling a story about Robin, the name of our robot, using a makeshift picture book. In this story we explained the similarities and dissimilarities between the robot and children to construct the type of common ground considered to have a positive effect on the learning outcome (Kanda et al., 2004). For example, we told that Robin enjoys dancing and wants to meet new friends, and even though he does not have a mouth and because of that cannot smile, he can smile using his eye LEDs.

After this story, another experimenter entered the room with the robot while it was actively looking at faces to provide an animate feeling. The robot introduced itself with a small story about itself and by performing a dance in which the children were encouraged to participate. The end of the circle time consisted of getting a blanket for the robot so it could “sleep”. This introduction was repeated later on the days we conducted the experiment in one-on-one sessions. While by then most children were comfortable interacting with the robot, some were still timid and anxious. To encourage these children to feel comfortable, one of the experiment leaders would sit next to the child during the warm-up phase of the experiment and motivate the child to respond to the robot when necessary until the child was sufficiently comfortable to interact with the robot by herself/himself. We found that the younger 3-year olds required more support from the experimenters than the older 3-year olds (Baxter et al., 2017). Although we are still analyzing the experiments, preliminary findings suggest that our introduction helped children to build trust and common ground with the robot effectively.

Research has shown that it is crucial for children’s language development that their communication bids are responded to in a temporally contingent manner (Bornstein et al., 2008; McGillion et al., 2013). This, however, faces a technological challenge. While adults tend to take over turns very rapidly, robots require relatively long processing time to produce a response. Nevertheless, in our first experiment (de Haas et al., 2016), we observed that children were at first surprised by the delayed responses, but quickly adapted to the robot and waited patiently for a response. Perhaps this is because children also require longer than adults to take turns (Garvey and Berninger, 1981) and having framed the robot as a peer children made the delays more plausible or expected. Nevertheless, while a lag in temporal contingency may not harm the interaction with children, it may harm learning. One way to remedy this may be to have the robot start responding by providing a backchannel signal, such as “uhm” to indicate the robot is (still) taking his turn, but requires more time to process (Clark, 1996).

Robots should not only respond to children in a timely fashion, but also in a semantically contingent fashion (i.e., consistent with the child’s focus of attention), as this too has a positive effect on children’s language acquisition (Bornstein et al., 2008; McGillion et al., 2013). For instance, research has shown that by responding in a semantically contingent manner, either verbally or by following children’s gaze, (joint) attention is sustained for a longer duration (Yu and Smith, 2016), allowing children to learn more about a situation. To achieve semantically contingent responses, the robot should be able to understand the child’s communication bids, construct joint attention with the child, or at least identify what the child is attending to. Monitoring children’s behavior and establishing joint attention are therefore considered crucial for designing a successful robot tutor.

To understand children’s communication bids, as well as to test their pronunciation of the L2, it is important that the robot be equipped with well-functioning automatic speech recognition (ASR). However, the performance of state-of-the-art ASR for children is still suboptimal, especially for preschool-aged children (Fringi et al., 2015; Kennedy et al., 2017). Reasons for this include that children’s pronunciation is often flawed and that their speech has a different pitch than adults. Moreover, relatively little research has been carried out in this domain and not much data exist to train ASR on. While it can be expected that the performance of ASR for children will improve in the not too distant future (Liao et al., 2015), until then alternative strategies need to be developed that do not (exclusively) rely on ASR.

In our project, we explore various strategies to achieve this, both based on monitoring non-verbal behaviors of the children and focusing on comprehending rather than producing L2. The first strategy relies on providing children tasks they have to perform in the learning environment, such as placing “a toy cow behind a tree” when teaching spatial language. This, however, requires the visual object recognition on the robot to work well, which is only the case when the scene contains a limited set of distinctively recognizable objects, such as distinctly colored objects (Nguyen et al., 2015). A potential solution explored in our project is to use objects with build-in RFID sensors that can be tracked automatically. The second solution we explore is to use a touch screen tablet that displays scenes the child can manipulate, which not only has the advantage of avoiding the problem of object recognition, but also allows us to control the robot’s responses and vary the scenes in real time. A downside, however, is that it takes away the 3-dimensional physical aspect of embodied cognition that would help the children to better entrench what they learn (Glenberg, 2010). Currently, experiments are underway to investigate the effect of using real vs. virtual objects. These solutions not only aid in understanding the child’s communication bids, it also helps in identifying their attention and can thus contribute to establishing joint attention.

Joint attention, where interlocutors attend on the same referent, is a form of social interaction that has been shown to support children’s language learning (Tomasello and Farrar, 1986). One way to establish joint attention with a child is to guide their attention to a referent using gestures, such as pointing or iconic gestures. The ability to produce gestures in the real world is potentially one of the main advantages of using physical robots as opposed to virtual agents, who may have a harder time to establish joint attention. However, many robots’ physical morphologies do not correspond one-to-one to the human body. Hence, many human gestures cannot be translated directly to robot gestures. For instance, the NAO robot that we use in our research has a hand with three fingers that cannot be controlled independently, so index finger pointing cannot be achieved (see Figure 1). Will children still recognize NAO’s arm extension as a pointing gesture? And if so, will they be able to identify the object the robot refers to? We are currently running an experiment to investigate how NAO’s pointing gestures are perceived, and preliminary findings show that participants have difficulty identifying the referred object on a small tablet screen. Similar issues arise when developing other gestures. One of the other non-verbal behaviors we are using is the coloring of NAO’s eye LEDSs to indicate the robot’s happiness as a form of positive feedback, since the robot cannot smile with its mouth.

Feedback, too, is an interactional feature known to help language learning (Matthews et al., 2007; Ateş-Şen and Küntay, 2015). The question is how should the robot provide feedback, such that it is both pleasant and effective for learning? While adults provide positive feedback explicitly, they usually provide negative feedback implicitly by reformulating children’s errors in the correct form. In child-child interactions, however, Long (2006) found that there was a clear advantage in learning from explicit negative feedback (e.g., by saying “no, that’s wrong, you need to say ‘he ran”’) when compared to reformulating feedback (the learner says “he runned” and the teacher reacts with “he ran”).

To investigate how children experience feedback from a peer robot, we carried out an experiment among 85 3-year-old Dutch-speaking children at preschools in Netherlands (de Haas et al., 2016, 2017). In this experiment, the children interacted with a NAO robot during which they received a short lesson on how to count from 1 to 4 in English. After a short training phase, in which the children were presented with the four counting words twice in relation to body parts and wooden blocks, they were given instructions by the robot to pick up a given number of blocks. While the instructions were given in their native language, the numbers were uttered in English. In response to the child’s ability to achieve the task, the robot provided feedback. The experiment followed a between-subjects design with three conditions: adult-like feedback (explicit positive and implicit negative), peer-like feedback (no positive and explicit negative) and no feedback. We did not find significant differences in learning gain between the conditions, probably because the target words were insufficiently often repeated. However, we explored the way in which the children engaged with the robot after they received feedback and we found that children looked less often at the experimenter in the feedback conditions than in the no feedback condition. Further analyses are carried out to evaluate how the children responded to the various forms of feedback to find out what type of feedback would be most effective for achieving both acceptable and effective tutoring interactions.

Finally, from a pedagogical point of view it is desirable that the interactions between child and robot be sufficiently challenging and varied so that the child has a target to learn from, but at the same time interactions should not be too difficult, because that may frustrate the child causing it to lose interest in the robot (Charisi et al., 2016). In other words, the robot should remain in Vygotsky’s Zone of Proximity that supports an effective learning environment (Vygotsky, 1978). In order to achieve this, the robot should be able to keep track of the children’s advancements in language learning and perhaps their emotional states during the tutoring sessions, and adapt to these. While the former can be monitored as discussed previously, it may be possible to detect emotional states known to influence learning (e.g., concentration, confusion, frustration and boredom) using methods from affective computing (D’Mello and Graesser, 2012). Using this type of information, it is possible to adapt the tutoring sessions by either reducing or increasing the number of repetitions, and/or change the subject (Schodde et al., 2017).