Children’s creativity–their ability to generate novel, surprising, and valuable ideas–is known to contribute to their learning outcomes, personal growth, and well-being. Creativity facilitates children’s problem solving, adaptability, self-expression and health (Carterette et al., 1994). Even though the benefits of creativity are widely recognized, classrooms are not able to sufficiently support children’s creative development. Gardner and Art, 1982 posited that children start to show creative abilities as early as preschool, and Smith and Carlsson (1983) found that the level of developmental maturity necessary for creative expression occurs around 5–6 yr of age. However, as they enter elementary school, children’s creativity slumps, especially around the fourth grade (Torrance, 1966, 1968; Claxton, 2005). As school curricula become more structured, children lose the aspect of creative play that is a significant part of kindergarten. To be successful in our AI-powered world, where mechanical and repetitive jobs are becoming automated, we must empower children to generate new artifacts and solve complex problems, which will require imaginative and novel thought.

In classrooms, social interaction plays a key role in children’s creative growth. Children learn creativity from teachers and peers who act as models for creative expression. They can scaffold children’s creativity through social interactions such as collaborating, posing challenges, asking questions, providing positive reinforcement, and generating ideas. Participating in collaborative tasks is one of the most effective external influencers of creativity (Kafai et al., 1995). In addition, classrooms today see increasing numbers of digital pedagogical tools and learning aids that have proven to be beneficial for cognitive learning due to their ability to personalize instruction for every student. Games and play-based learning approaches have been successful in fostering creative expression in children (Henriksen, 2006; Bowman et al., 2015). However, most educational technologies do not have the ability to foster social interaction with students. Exceptions include socially interactive AI agents such as conversational agents and social robots, which are highly effective in promoting learning and engagement (Chen et al., 2020). Previous work has demonstrated how social robots can influence children’s learning behaviors such as curiosity, growth mindset, and empathy through social emulation (Gordon et al., 2015; Park et al., 2017). Notably, the presence of a social robot can also affect adults’ creativity (Kahn et al., 2016; Alves-Oliveira et al., 2019). In this work, we explore how social robots can foster creativity in young children through collaborative, playful interactions.

Social robots are increasingly being used as learning peers and tutors (Adamson et al., 2021), and their unique ability to socially interact with children while being co-located has situated them well as creativity support tools (CST). Previous work has shown that robots are more effective than other mediums used for CSTs, such as screen-based interfaces (Kahn et al., 2016; Ali, 2019). We do not seek to compare robots to other mediums of CSTs in this work; rather, we intend to demonstrate the efficacy of creativity stimulating interactions designed for social robots. Furthermore, while social interaction is not a prerequisite for creativity, creativity literature informs us that social interactions with peers and tutors can foster creativity in children. Previous work has shown how situating the robot as a collaborative peer that offers ideas or helps with the creative process have benefitted creative expression (Louis and Peter, 2015; Rond et al., 2019). In this work, we explore whether a social robot’s capability for social interaction patterns can stimulate children’s creativity, while acknowledging that there are other stimulants of creativity. Learning from the effect of sociality on creativity in classrooms, we explore whether the effect can be replicated in pedagogical tools, specifically social robots. We suggest two interaction patterns in which intelligent embodied agents can help children think more creatively: 1) creativity modeling, where the social robot models or demonstrates desired creative behaviors, and 2) creativity scaffolding, where the robot offers scaffolding to the child in the form of asking reflective questions, validating novel ideas, and engaging in creative conflict. The robot used in this work is Jibo -- a child-friendly, tabletop, socially expressive robot.

We position our research in three playful and collaborative tasks, where children and the robot collaborate to create artifacts. These one-on-one interactions afford different forms of creative expression.

In the first two tasks, outlined in our previous research (Ali et al., 2019; Ali et al., 2020), we designed the behavior of the robot to artificially emulate human creativity:

1. The Droodle Creativity game, where children and the robot generate humorous titles for abstract images to express verbal creativity.

In this work, we introduce the third task, where the robot scaffolds children’s creative thinking by asking questions, validating novel ideas, and engaging in creative conflict:

3. The WeDo Construction activity, where children and the robot co-create WeDo LEGO models to encourage constructional creativity.

We outline previous research, where we demonstrated that children can emulate a social robotic peer’s creative expression during collaborative gameplay. We ran two randomized controlled trials with 126 children: verbal creativity (n = 48) and figural creativity (n = 78) in the 5–10 yr old age group. Participants in the intervention group interacted with the robot exhibiting creative behaviors and participants in the control group interacted with the robot that did not exhibit these behaviors. We observed that children who interacted with the robot exhibiting high verbal creativity (in the Droodle Creativity game) and high figural creativity (in the MagicDraw game) exhibited higher verbal and figural creativity themselves. We then introduce our third study, where the robot offers creativity scaffolding behaviors in the WeDo Construction task. We ran a randomized controlled trial with 42 students in the 6–10 yr old age group. We observed that when the robot offered creativity scaffolding in the construction task, children expressed higher creativity.

In sum, we provide consistent evidence that the performance of creativity inducing behaviors by social robots can foster creativity in young children. Further, in all three studies, children in the high creative robot conditions perceived the robot as more creative and fun as compared to the low creative conditions. By showing that a social robot can successfully foster different kinds of creative expression, we are able to articulate more generalized social mechanisms that can be leveraged to support creativity in children. We contribute novel design guidelines and new methods for designing interactions with social agents that aim to promote creative thinking. With new developments in generative modeling techniques, robots can participate in several co-creative tasks with children, and can be leveraged as creativity support tools in a wide range of creative activities. We discuss implications for the field of HRI, digital creativity support tools, co-creative agents, and transformative games.

For the three creativity activities, we used Jibo, a socially embodied robot, as our robotic platform (Jibo, 2015). Jibo is an expressive tabletop social robot (Figure 1) that can speak, respond to children’s speech, track faces, gaze toward the person, attend to sound and movement in its environment, and physically express emotion through its display and three degree-of-freedom body. Jibo can communicate with an Android tablet that serves as a shared drawing surface or to display information for the child. Prior to each study, the experimenters discussed with children that Jibo uses WiFi to see, talk, draw, and interact with objects displayed on the tablet, and that it doesn’t need physical hands to do so. Such discussion is important to building a believable experience for children that the robot knows what is displayed on the tablet and can draw on the tablet, too.

During the interaction, Jibo provides explanations and encouragement to the child, and expresses joy, curiosity, and pride. Between the high creative (C+) and low creative (C−) robot conditions, we carefully controlled the amount of verbal and non-verbal robot behaviors; however, the robot’s speech differed based on the study condition. In the C+ condition, the robot exhibited greater creativity; in the C− condition, it demonstrated substantially lower creativity.

We developed three game-based and interaction child-robot tasks which we introduce in the following sections.

Previous work in HRI demonstrated how children emulate robots’ learning behaviors such as curiosity and growth mindset (Gordon et al., 2015; Park et al., 2017). Motivated from this literature, we explored whether this social emulation phenomenon extends to verbal creativity. Verbal creativity, defined as the ability to create verbal artifacts such as stories, prose and poetry has three indicators: fluency, or the ability to produce a large number of ideas; flexibility, or the aptitude for changing from one approach to another or from one line of thinking to another; and originality, or the capacity for bringing new ideas or solutions that are far from obvious, common, or established. Development of verbal creativity in children is pivotal to their learning, writing and thinking skills and helps them reflect their feelings, emotions, opinions, reactions, and notions to others (Shorofat, 2007; Rababah et al., 2017). In this work we explore whether a social robot’s creative verbal expression is emulated by children, for which we created the Droodle Creativity game (Ali et al., 2019), inspired by the Droodle Creativity Task, a verbal creativity task that draws upon people’s ability to creatively use language to describe an abstract image or figure known as a droodle (Kahn et al., 2005). The Droodle Creativity Task has been previously validated as a means to measure people’s verbal creativity, and is based on the cartoon book Droodles by Roger Price (1982), thus making it appropriate for a children’s game. The Droodle game encourages children to think creatively, express their thoughts and encourage humor.

Similar to verbal creativity, we aimed to explore whether children also emulate a robots’ expressed figural creativity. Generative models such as GANs make it possible for AI models to generate creative drawings (Ha and Eck, 2017; Ge et al., 2020). Collaboration with both humans and digital interfaces has also been known to benefit figural creativity (Kim et al., 2016). Collaborative digital CSTs have been constantly evolving to support the creativity needs of people, and have the ability to contribute to the drawing itself. In human-human creative collaborations, creators can socially interact with one another, provide feedback and comment on their drawings, an interaction style that is lost in digital CSTs. In this work, we explore whether collaborating with a robot that also interacts with the creators socially, in a figural co-creation activity benefits children’s creativity. We explore whether children emulate the robot’s expressed figural creativity through a co-doodling task.

Physical construction, or the ability to make new artifacts by combining other artifacts, is a key indicator of children’s creativity. Furthermore, unique ways of combining and using the creations also indicates divergent thinking. Physical construction kits facilitate playful learning, open-ended making and creativity, and often creates co-creation space with others (Alimisis, 2013). Robotic toolkits and environments have been successfully leveraged to afford construction by children (Mioduser and Levy, 2010). When children construct with others, their creativity is scaffolded by other children acting as models to emulate, providing ideas, brainstorming, challenging and asking questions. In this experiment, we explore whether this social support can be offered by a social robotic collaborator in a construction activity.

To afford constructional creative expression through making, we designed a third activity in which children and Jibo collaborate building models using the LEGO Education WeDo 2.0 Core Set (2020). The set consists of LEGO bricks and electronics that can be programmed using a visual programming interface on a tablet application with the aim of introducing children to computational thinking and engineering principles in a fun and engaging way.

The interaction involved children making construction projects using the WeDo 2.0 kit in the presence of a social robot, which assumed the role of a tutor and provided scaffolding to the child. In order to get acquainted with the programming interface, children were first guided by the robot to build a rover using LEGO blocks, and to program the rover such that it could detect obstacles and respond to their commands using the WeDo Android tablet application. Children could utilize WeDo 2.0 standard construction kit items, including a Bluetooth enabled controller, LEGO bricks, motors and supporting construction materials and motion sensors. This guided activity was conducted through a step-by-step verbal exchange between the child and the robot and lasted for 6 min, with the robot taking the instructor role (Figure 5). The activity introduced the child to sequential commands, condition statements, delays, and loops. Then, children were given 20 min for free play, where they could explore different functions of the WeDo app, add new LEGO blocks, and make their models perform new actions. The idea generation process was guided by both the child and the robot. The role of the robot was to scaffold the child’s creative learning through verbal and non-verbal behaviors. Throughout the interaction, children could ask the robot questions and receive dynamic troubleshooting guidance. The robot also engaged in active creativity scaffolding which involved asking the child reflective questions, challenging their ideas and assumptions, and suggesting alternate ideas for creations with the rover. The robot also provided feedback and positive affirmation after children generated new ideas.

While the Droodle Creativity game and MagicDraw game utilized a fully autonomous Jibo interaction, in this activity, the robot was controlled by a human instructor using a dynamic and predictive Graphical User Interface (GUI) on the desktop to provide quality scaffolding at the right time and also to respond to the child appropriately. This GUI controller was iteratively co-designed in tandem with the instructors, with the goal of assisting them in providing creativity scaffolding to the children. This iterative process of designing the GUI is outlined in Section 3.4.2. The desktop GUI application communicated with the robot using Robot Operating System (ROS, 2021). Children programmed the WeDo controllers using an Android application on a tablet screen. Figure 5 illustrates the system components and communications between them.

In this work, we demonstrate how a social robotic peer’s socio-behavioral patterns can influence creativity in children in the 5–10 yr old age group. Specifically, we highlight two robot interaction patterns: 1) creativity demonstration, where the robot itself demonstrates artificial creativity, and 2) creativity scaffolding, where the Jibo robot supports and encourages the child’s creative thinking by asking reflective questions, providing challenges and positive reinforcements. We designed three game-based interactions that afforded different types of creativity. The robot’s interaction patterns were inspired by children’s creativity-eliciting social interactions with their peers and tutors. These interactions serve as playful ways of measuring creativity, as well as supporting children’s creative expression. In order to assess the efficacy of these interaction patterns, we conducted an accompanying investigative study for each of our game-based interactions: the Droodle Creativity game that affords children’s verbal creativity, the MagicDraw game that affords children’s figural creativity, and the WeDo construction task that affords children’s constructional creativity.

In the Droodle Creativity game, children emulated artificial verbal creativity exhibited by Jibo during gameplay. Participants who interacted with the high creative robot expressed more ideas, more diverse ideas, and highly creative ideas in the Droodle Creativity game, as compared to participants that interacted with the low creative robot. Similarly, in the Magic Draw Game, children emulated artificial figural creativity exhibited by Jibo in a co-drawing task, which led to their drawings being measured as more creative by the TCT-DP test for figural creativity. Participants also perceived the high creative robot as highly creative. Through these two studies, we verify our first hypothesis: that children adeptly socially emulate the creative behavior of peer-like robot playmates, and this in turn fosters children’s own creative behavior. Importantly, this is a sufficiently robust enough finding that we could replicate it for two different kinds of creative expression: verbal and figural creativity.

In the second investigation, we demonstrated how a robot offering creativity scaffolding in the form of asking reflective questions, challenging the participants, and providing positive reinforcement had a positive effect on children’s creativity. Participants engaged in an open-ended activity involving constructing and programming a rover using the WeDo construction kit. Creativity was measured by the ideas that children came up with for the rover, the number of different tools (programming bricks) they used, and how unique their ideas were. Children interacting with the robot that offered creativity scaffolding scored significantly higher on the number of different ideas and different programming bricks used. They also scored higher in the uniqueness of their ideas; however, that difference was not statistically significant. Hence, we could establish evidence that children can learn creativity from a social agent by emulating the agent’s creative behaviors and by the agent scaffolding their creative learning, which informs the design of pedagogical embodied tools to foster creativity.

The WeDo construction task utilized a predictive scaffolding model. This paves the way toward the development of fully autonomous robot scaffolding systems for tailoring personalized learning to different students and contexts. Over time and over several playtests, the model is able to reinforce itself depending on whether the instructor accepts or rejects its suggestions, eventually leading to minimal error rates. Replicating a human instructors’ scaffolding into an artificial agent can be beneficial for personalized assistance when the teacher is not present or when there are many students per teacher. This scaffolding paradigm could be used in the context of other activities.

It is important to be wary of the shortcomings of such a suggestion based system for instructors. While these recommendations make it easy for instructors to provide help and scaffold children’s creative learning, they also inhibit the instructor’s original thought and manner of scaffolding, which holds high value. A more autonomous model would be built from data collected from multiple instructors with a diverse set of backgrounds and expertise, all of whom could instruct several students from a diverse set of backgrounds. Further, the model should be able to adapt based on an instructor’s usage while allowing enough space for original thought. While the current interface does allow instructors to reject the model’s recommendation and instead generate new Jibo utterances, it can still influence the instructors’ decision making process, and cause them to conform to commonly used instructions, which may be counterintuitive to promoting creativity. In our work we suggest two interaction patterns of social robots that effectively foster creativity in young children: creativity demonstration and creativity scaffolding using questions, questions and ideations. We position social robots as CSTs in collaborative activities with children that can leverage the benefits of having a co-located peer to create and socially interact with, and a digital CST that adapts to the person’s creation style. We also add to the literature of HRI suggesting that children emulate a robot’s learning behavior and extend it to verbal, figural and constructional creativity. Through our observations from playtests, iterative game design, study results and post-test questionnaires, we formulated evidence-based design guidelines as well as additional recommendations from researchers’ reflections for designing creativity supporting social agents that we have outlined in the next section. These recommendations could benefit pedagogical researchers, educators and HRI and HCI practitioners designing social agents for stimulating creativity, especially in children.

In this work, we posit social robots as CSTs for children in collaborative tasks. We studied the effects of an autonomous social robot’s verbal and nonverbal interactions on children’s creativity as measured by three collaborative game-based child-robot interactions. We observed that both creativity demonstration and creativity scaffolding offered by the social robot had a positive effect on children’s creativity in verbal, figural and constructional creativity tasks. This work contributes to the design of game-based child-robot interactions that afford creativity, provides evidence for the efficacy of these interactions and provides guidelines for designing social embodied agents to foster creativity in young children. These findings are valuable to game designers creating game-based interactions to foster children’s creativity, as well as HRI and HCI practitioners leveraging social agents as CSTs.

Since robots are already being used in classroom settings as learning peers and personalized tutors, it is imperative to think about how their behaviors can influence children’s learning behaviors, such as creative thinking. While social robots are not the only way to provide creativity support through behavioral modeling, they certainly are a compelling way given their social nature. Effort must go into designing the agents’ behavior such that they exhibit creativity and scaffold the child’s creativity as a peer or a tutor. Embodied AI agents have the potential to use generative modeling techniques to express different forms of creativity through generating media such as drawing, poetry, art styles, patterns, physical body movements, etc. They are also socially emotive and can express the social interactions that accompany creativity such as reflection, inquisitiveness and positive affect. This work opens up opportunities to explore how these different forms of artificial creativity can be embedded into tools that children use and interact with, and help them be more creative.

Recent works have demonstrated how robots can help creativity, as a co-present partner and through social interactions. In this work, we suggest two interaction patterns of social robots that we observed to effectively foster creativity in young children: creativity demonstration and creativity scaffolding using questions, questions and ideations. We add to the literature of HRI suggesting that children emulate robots’ learning behaviors, and that this phenomenon extends to creativity. We also contribute to the field of Creativity Support Tools by positioning social peer-like robots as a creativity support peer in collaborative activities. This contribution is not only valuable for HRI practitioners, but also other interactive AI agents, such as conversational agents. Creativity supporting social robots combine the benefits of having a co-located peer to collaboratively create with and socially interact with, and a digital CST that adapts to the user’s creation style. Leveraging generative AI models now allow for expressing creativity in several modalities, which robots can successively leverage. Further, we elaborate the design of a scaffolding mechanism by learning from human scaffolders through one construction activity. This approach can be generalized to other activities.

While introducing an extrinsic factor in the form of a social robot, we must ensure that it does not come across as an evaluator; classroom research has demonstrated how extrinsic factors such as evaluation, competition and unrealistic expectations can potentially inhibit creativity, instead of fostering it (Torrance, 1967). In the scaffolding GUI design, we provide instructors with a predictive interface that helps them scaffold the child for creative learning; however, this suggestion model also limits creativity and personalization of teaching style from the instructors. To tackle this issue, we must also aim to build personalized scaffolding models that take input from every teacher and personalize over time in both the content and style of learning.

In our work, we chose a wide age range (5–10 yr) and we did not analyze differences across age ranges. This is a limitation of the current work, and future analysis needs to be run with narrower age ranges to determine the efficacy of the intervention on different age groups. Another limitation of this work is that these robot interactions lead to an increase in creativity within the narrow constructs of these tasks, and may not scale to every creative task, or to students’ life outside of these tasks. There are also countless ways of expressing creativity such as poetry, storytelling, painting, music, etc., that are not explored in these tasks. This work also defines a limited scope of creativity in terms of fluency, novelty and value of ideation. Creativity encompasses a much wider array of behaviors (such as divergent thinking) that can be explored using other interactions. Furthermore, while all these interactions currently focus on one-on-one child-robot interaction, we must strive towards designing interactions that involve multiple children because collaboration with peers forms a major part of creative learning.

Finally, while this work evaluates the role of the robot’s creativity fostering behaviors, it does not evaluate the benefits of embodiment over other non-embodied agents such as computers or voice agents. Previous research has found that adults did not show significant gains in creativity merely in the presence of a social robot (Alves-Oliveira et al., 2019). In future work, we aim to study the combined effect of embodiment and creativity scaffolding behaviors by running a 2 × 2 study ({embodied, non-embodied}×{scaffolding, non-scaffolding}) (Devasia et al., 2020). In order to evaluate whether robots are really that social, future work is required to assess the creativity effects of human peers vs robots.

While this work makes use of Jibo as the social agent, these interactions are learning tools that can foster creativity in classrooms and homes independent of Jibo. These games also serve as game-based creativity assessment measures.

We designed a creativity scaffolding paradigm for the WeDo construction task. This model currently supports a semi-autonomous scaffolding system, where a human controls the robot using a remote control desktop program. In the current version of the robot control interface, which lets instructors control the robot remotely, we can incorporate ASR to use the instructors’ speech to control the robot’s speech. Moreover, collecting more data about how instructors use the program can help us build a fully autonomous model of scaffolding. While this approach can be used to design interactions for an autonomous or semi-autonomous system, the timing of the interactions and during-study improvisation are reliant on children’s actions in the task and interactions with the robot. Current status of natural language understanding and computer vision limit a complex understanding on the scene, and hence a similar fidelity of robotic interactions are challenging to currently implement in a fully autonomous system. However, as we demonstrated, a semi autonomous system (where the system suggests interactions and the human decides the timing of the prompts) is feasible.

Another limitation of this work is that all activities are self-contained and involve single interactions. In future work, it would be valuable to evaluate creativity transfer from one activity to another, and even in the absence of the robot in the long-term. Design recommendations from this work can be incorporated in several creativity support tools such as computer games, voice agents, tablet apps, embodied tools, space design, etc. Finally, advances in generative modeling techniques enable us to create child-robot interactions supporting multiple modalities of autonomous creative expression.