Sinem Aslan (Ph.D.) is a Senior Research Scientist at Intel Labs, focused on improving the capability of AI to understand and model human behavior for enabling personalized experiences. Lenitra Durham (Ph.D.) is a Senior Software Research Engineer/Scientist at Intel Labs, architecting and prototyping solutions combining UX research with AI and validating the solutions in real-world environments.
Oscar may be a digital character in Intel Labs’ Kid Space research prototype, but it’s the kids who become animated in the presence of this friendly, AI-enabled bear (see Figure 1). Oscar invites them into his world of collaborative play-based learning, and soon they are dancing, laughing, and engaged with one another as they work to help him accomplish various math-based tasks. This is just one of the key observations from Kid Space research pilots deployed between 2020 and 2022. In this blog, we'd like to briefly introduce our Kid Space research and discuss some of the user insights we’ve gained from these pilots so far.
Figure 1. Kid Space Peer Learner, Oscar the bear, asks children to help him solve mathematical-based problems.
What Is Kid Space?
Parents of young children share many of the same concerns and beliefs when it comes to technology-supported learning. Our ethnographic research with dozens of parents of young children showed that most of the parents recognized the usefulness of technology in learning but worried that too much screen time would result in insufficient social engagement and physical activity. To address these pain points, we set out to design a fully immersive math learning experience that would engage both the minds and bodies of young learners through pedagogically appropriate technology use.
Kid Space incorporates audio and vision hardware components and AI technologies to create an adaptive and engaging learning environment (see Figure 2). Oscar, our peer learner, uses multimodal sensing and sense-making technologies to understand the physical cues from children and respond with personalized learning opportunities. Key features of the AI system are (1) multimodal interaction using voice, touch, and vision-based cues (e.g., pose, gesture, etc.), (2) object segmentation, recognition, and tracking to enable the use of physical manipulatives in learning, while creating a mixed-reality experience, and (3) user tracking to help Oscar personalize the spoken dialogue interactions with the children.
Figure 2. Kid Space interaction space.
Table 1 shows the evolution of our Kid Space research Proof of Concept (POC) and the user research pilot studies that we have conducted in schools and homes to date.[i]
Table 1. Kid Space Research Timeline Overview
Positive Educational Outcomes from Kid Space in Schools
Over the last two years, our Kid Space prototype for schools has evolved from a “Wizard of Oz” type POC, which relied heavily on behind-the-scenes human control of Oscar, to our first end-to-end research POC utilizing human-AI collaboration. We used our POC in a recently completed research pilot to evaluate its effectiveness and understand experiences and perceptions of participants. We tested the POC at an elementary school in Hillsboro, Oregon with fourteen first-grade students, two educators, and one technology coordinator (see Figure 3).
Figure 3. Sample pilot view. Children engaging with Kid Space designed for school usage.
While we are still in the process of analyzing the data and finalizing the user insights, we are excited with the following preliminary findings:
- Improved student performance. The children's scores on a post-math test implemented after the sessions were significantly higher than their scores on a pre-test implemented before the sessions (p<0.05).
- High levels of student engagement. Coarse-level observations showed that students remained on-task and satisfied (>90% of the time) during the sessions. Students also demonstrated excitement—a rarely observed emotion in learning contexts.
- Improved student positivity. The students’ ratings of how they felt after Kid Space sessions were significantly more positive than their ratings before the sessions (p<0.05). In addition, all fourteen students said they wanted to play with Oscar again.
- Positive social interactions with peers and Oscar. Coarse-level observations showed that the students demonstrated positive social interactions with their peers and Oscar during the sessions.
- Positive perceptions toward Oscar and Kid Space. The three school personnel who observed multiple sessions gave a high rating to Oscar as a conversational agent (with an average score of 9.2, on a scale of 1-10), and Kid Space design principles on a scale of 1-5 as follows: Educational (5), social (4.7), magical (5), creative (4.7), physical (5), and renewable (4.7).
Home-Based Online Learning: What Teachers Want
The COVID-19 pandemic spotlighted the many limitations of online learning, leaving teachers and parents frustrated and overwhelmed, and young learners underserved. It also amplified widely acknowledged concerns that excessive screen time—and the associated lack of physical activity and traditional play—could negatively impact children's physical, emotional, and social development. Shortly after the shutdown ended, we initiated another study to improve our understanding of how online learning might be improved. In this study, we interviewed public and private school teachers about how they implemented online learning during the shutdown, the challenges they faced, and the suggestions they had to address these challenges. Among other things, we asked the teachers what they most desired in home learning technology. A summary of key insights gained from their answers is shown in Table 2.
Table 2. Teachers’ Wish List of Technologies to Address Online Learning Challenges. For more details about this study, please see our recently published journal article.
Using the key insights derived from this study, we created a new home-based Kid Space research POC, which we tested in the homes of twelve second-grade students and their parents. We used a Wizard of Oz approach to iteratively test the designed experiences. Figure 4 shows the basic components of the system including a laptop with the conversational agent, a built-in microphone and camera on the laptop for multimodal interaction, a play mat with tangible math manipulatives, and a depth camera for manipulatives detection, recognition, and tracking. Figure 5 shows a sample pilot view of Kid Space at home.
Figure 4. Setup overview of home-based Kid Space POC.
Figure 5. Sample pilot view. Child engaging with Kid Space designed for home usage.
Our data analysis yielded the following key insights:
- High levels of engagement, physical activity, and social interactions. 11 out of 12 children and their parents rated Kid Space highly in terms of engagement, physical activity, and social interaction – especially when compared to traditional PC-based educational games.
- On task and satisfied. The children demonstrated on-task behavior 100% of the time and satisfied behavior approximately 80% of the time. The children also demonstrated excitement (a rare emotion to observe in a learning context) approximately 20% of the time.
- High levels of engagement with less screen time. Although the children were highly engaged with the Kid Space games, their attention was split fairly evenly between the screen and physical elemental aspects of the experiences (e.g., manipulatives), with overall screen time dropping as low as 50%. After observing the Kid Space session, the parents’ concerns about screen time were significantly lower than their concerns about their children’s usage of traditional PC-based games. (p<0.05).
- High physical activity. The children were physically active about 90% of the time when engaged with Kid Space—a vast improvement over the sedentary nature of traditional PC-based educational games.
What’s Next for Kid Space?
These promising initial findings suggest multimodal conversational AI could be effectively used to improve educational outcomes and experiences for young children, teachers, and parents. We will continue to explore possibilities with an improved Kid Space research POC that includes algorithms for resilient recognition, additional scaffolding learning strategies, more opportunities for personalization, and other improvements as derived from the user research. We will also conduct further research to define an educational platform that allows stakeholders to add and customize content in Kid Space to address scalability and decrease the burden on developers when designing such AI systems.
We would like to emphasize that our Kid Space research is rooted in social science and that our multi-disciplinary research team works with students, teachers, parents, and other stakeholders in the field to maintain responsible AI practices and technologies in our educational research. As researchers in this domain, we are cognizant of the many potential areas of concern including misuse, bias (especially with respect to neurodiversity), lack of equity in access, robustness, privacy, security risks, and more. We assess these risks in our development and deployment lifecycles and research ways to mitigate them; we will continue to leverage what we learn. We look forward to future research pilots as we work responsibly and collaboratively with stakeholders to find ways to improve educational outcomes with fully immersive learning experiences.
For more information about our Kid Space research, please see:
Building Wonderful: Human-AI Collaboration
AI for Social Good – Intel Responsible AI
[i] Please note that prior to these research pilots, we went through Intel’s rigorous Privacy Impact Assessment process. We also sought and received Intel Legal’s approval to conduct user research with parents and their children, as well as the educators. All research participants signed consent forms prior to the studies informing them of all important details about the studies, including the research goals and procedures, as well as how their data would be collected and used to support Kid Space research. As with all research projects, Intel and its collaborators abide by strict data privacy policies and adhere to ongoing oversight. For more information, please refer to: Intel Privacy Notice.
Sinem Aslan earned her M.S. and Ph.D. degrees in Instructional Systems Technology with a minor in Educational Inquiry Methodology from Indiana University. She currently works as a senior research scientist at Intel Corporation. Her research focuses on understanding how AI can better understand and model human behavior to enable personalized experiences. As a scholar, she authored several journal publications, given presentations at international conferences, and received a number of internationally-recognized awards for her research.