The goals of this Expedition were to advance the computational science in orderto accelerate the realization of social robots capable of promoting the education and health of young children over a range of populations. The core intellectual merit of this work is the successful advancement of embodied computational intelligence -- including the development of novel algorithms, methodologies, metrics, and validated prototypes of socially assistive robot systems that successfully demonstrated the ability to personalize social behavior and collaborative engagement to meet the learning needs of individual children, over weeks and months.

Key efforts focused on social robot system and algorithmic development, long-term deployment in schools, and assessment of intervention efficacy for pre-K children. The robots were deployed in Boston-area public schools, representing a diverse student population including English language learners as well as those with other special needs, to promote kindergarten readiness targeting oral language and early literacy skills. The robots were designed to engage young children as peer-like learning companions; they played educational games together using a tablet computer as a shared game surface.  A set of educational games and activities were developed, focusing on storytelling for oral language development with target vocabulary.

Over the course of the Expedition, two new child-friendly social robot platforms were designed and developed to be suitable for young children. The social robots were designed with rich multi-modal expressivity to convey a wide range social and emotional cues. They also had a friendly appearance and plush exterior. Based on interviews and rating scales, the robot designs were very appealing and engaging for young children. New computational models to support the dynamics of social interaction between social robots and with young children were developed to address key issues of sustaining engagement, building trust, and maintaining a collaborative alliance to work toward learning goals. A key challenge was to develop and train computational models to enable a social robot to convey attentiveness, engagement, and understanding through generating dynamically contingent and appropriate non-verbal backchannel feedback cues in response to children’s non-verbal cues as they told stories to the robot.  We trained our model using a novel multi-modal dataset we collected of children telling stories to each other, and we have made this dataset available to support the research community. We validated that this model successfully captured key social dynamics so that the robot could support children’s storytelling behavior as and engaged and active listener.

We conducted four long-term deployments of our social robot systems in public schools lasting 2-3 months each. These deployments were used to evaluate the performance of our personalization algorithms, as well as to assess the impact and efficacy on children’s engagement and learning outcomes. Randomized controlled trials were conducted comparing the impact of a personalized social robot “learning companion”, to one that followed a fixed curriculum, to classroom baseline with no robot intervention. Key engagement measures examined children’s sustained positive valence over repeated encounters, responses to the robot’s prompts, and child-robot relationship measures. Key learning outcomes focused on vocabulary acquisition and oral language syntax gains measured pre/post.

We applied reinforcement learning methods to learn a personalized intervention policy of robot actions for each child based on the child-learner’s state. The robot’s actions included expressive behaviors intended to motivate and encourage each child, pedagogical actions intended to explain or ask a question to encourage children to reflect on the educational game/activity, as well as to adjust the level the difficulty of the game/activity to appropriately challenge each student. Detailed analysis of the data showed that the personalized learning companion system resulted in stronger engagement, vocabulary acquisition, and oral language improvement over the non-personalized system, and that both social robot interventions led to improved learning gains over baseline.  

The broader impacts of this Expedition were manifest in three areas. The first is the successful proof-of-concept application of socially assistive robots to key elements of early childhood learning and Kindergarten readiness. The potential for effective, affordable, scalable, and personalized intervention in early childhood education is an area of significant societal interest. Second, the Expedition supported key outreach activities including the development of a novel social robot toolkit to support young children’s STEM education. This included hands-on workshops in classrooms and afterschool programs, guided by a mentor, whereby children could partake in a set of structured activities to build simple smartphone-enabled social robots, program them with a developmentally appropriate programming environment, and train their robot creations to interact with them. Through these guided activities, children were introduced to key concepts in AI and robotics. Finally, this Expedition provided significant opportunities for comprehensive training of graduate and undergraduate students in all aspects of this research program resulting in numerous peer-reviewed publications.

Last Modified: 05/31/2018
Modified by: Cynthia Breazeal