The paper presents a robotic architecture for managing behavior in educational settings, inspired by Applied Behavior Analysis (ABA). The system features a social robot that (1) learns a model of a child's goals, beliefs, and intentions, grounded in observations and conversations, to explain their detected behavior, and (2) responds efectively to behaviors by planning appropriate sequences of actions to implement the strategies suggested by experts. By leveraging cloud-based processing and local execution, the robot dynamically adapts to social interactions in real-time, delivering educational activities, monitoring children's behavior, and applying behavior management strategies. Results from the experimental evaluations highlight the system's replanning and cloud response times, as well as its overall efectiveness.
Section 2 introduces the ABA methods and describes the scenario. Section 3 details the system’s architecture and behavioral planning. Section 4 discusses a case study of behavior function recognition. Section 5 summarizes the results. Finally, Section 6 discusses the conclusions and limitations.
Applied Behavior Analysis is an evidence-based approach aimed at improving socially significant
behaviors [
FBA also provides valuable support to educators by ofering a structured method for managing complex behaviors in a classroom setting.
Given this premise, the problem we aim to address is how to develop an autonomous social robot capable of planning educational activities, continuously monitoring children’s actions, conducting FBA to identify the purposes of challenging behaviors, and proposing alternative strategies to handle them. Our goal is not to categorize behaviors but to address a common classroom struggle: managing behaviors considered “challenging” because they present major challenges to maintaining a productive learning environment. A robot that can tailor learning interactions and respond efectively to others’ behaviors through the use of efective strategies could greatly improve the integration of robotics in education.
To achieve efective and autonomous behavior management in CRI, we have designed a software architecture that integrates cloud-based processing with local execution capabilities. Figure 1 illustrates the detailed layout of the proposed system architecture, which is divided into two main components: the Server and the Client.
The server hosts two primary elements: the Hub and the Planner Engine. The Hub manages connections to the external cloud service OpenAI, while the Planner Engine leverages Fast Downward, a highly eficient planning system based on the Problem Domain Definition Language. The Planner Engine processes inputs from the client, specifically the PDDL domain and the initial problem. These ifles outline the available actions, their efects, preconditions, and the initial state of the problem that the robot needs to address.
The client, embedded within the robot, comes preloaded with the necessary domain and initial problem files. The Plan Manager is responsible for sending these files to the Planner Engine and retrieving the generated plan.
Upon receiving the plan, the Behaviour Manager takes charge, iterating through the plan and determining the actions that need to be executed by checking the predicates.
As actions are performed, the PDDL Predicate Retriever asserts new predicates about the current state of the world based on the interaction with the child and the environment. The goal is to monitor plan progress to ensure that the actual preconditions and efects of actions match the expected ones.
On one hand, the PDDL Predicate Retriever employs its sensors to make assertions about the current state of the world; on the other hand, it attempts to infer the child’s mental state through conversation. To this end, the robot employs Microsoft Azure services to transcribe the child’s spoken words captured by microphones. The transcribed text is then sent to OpenAI, which analyzes the content to generate appropriate responses and identify predicates related to the child’s behavior.
The Behavior Manager compares asserted predicates with the expected preconditions and efects in the plan. If the predicates do not match, the Behavior Manager triggers the Plan Manager, which may then initiate a replan. This process follows a match and go approach, where the robot continuously evaluates the environment against a set of predefined predicates and conditions that must be met for an action to be executed. When a new situation arises, the robot checks its current state against the efects of the current action and all other actions’ preconditions. If these conditions match the expected criteria, the robot can either proceed to execute the next action — hence the term go — or request a replan. This method ensures that the robot’s actions are contextually adaptive. 4. A case study: Escape and Gain a Tangible Drawing from ABA and FBA principles, two behavior functions were identified to be recognized for preliminary testing of the model’s validity: Gain a Tangible and Escape. The aim is that, through verbal interaction with the child, the robot would be able to autonomously recognize the occurrence of the following behaviors: • Gain a Tangible: When pleasant activities are interrupted by time limits, negative reactions of the child may be elicited. • Escape: When presented with a task, a child may perceive it as too dificult or too boring, potentially provoking adverse responses to avoid the task.
With a comprehensive understanding of these psychological approaches, we developed a PDDL domain capable of autonomously managing various scenarios of interest according to the ABA approach. Leveraging actions detailed in Table 1, the framework can adapt to diferent situations, including making real-time adjustments. The initial plan involves a robot presenting a memory card game to a child. The process begins with a brief interaction (Interact), followed by an explanation of the rules (PresentTask).The robot then monitors the child while they play (Wait). As time nears its end, the robot warns the child (WarningCleanUp), instructs them to stop and put away the cards (PutAway), confirms they have finished ( ConfirmAllGood), and ends with praise (GoodJob). During the Wait action, the robot may detect issues by querying students via OpenAI-supported verbal interactions.
The paragraphs below present some examples of verbal interaction logs related to this situation. User inputs are denoted by “U,” the robot’s responses by “R,” and the retrieved predicates by “P.”
P: warning, not_waiting, want_to_play U: No, don’t worry, I don’t want to play anymore.
R: Giulia, why are you so upset? P: warning, not_waiting, want_to_play, not_agreed, not_doing Replan is needed
In the exchange above, the robot detects a shift in the child’s engagement, recognizing that Giulia is no longer interested in continuing the game. This prompts replanning starting from the current state (see Section 3), with the asserted predicate not_doing acting as a trigger. In response to this interaction, a new plan is formulated, which includes the action (DiscoverToM1 game kid memory). The efect of this action is investigating, and the following log shows a possible outcome of the interaction.
P: investigating, task_is_hard Replan is needed
This update leads to the conclusion that the task is perceived as too challenging for the user, triggering again the PDDL Predicate Retriever to update the predicates describing the current state. It causes a mismatch with the expected efects of the current action that is detected by the Behavior Manager, which prompts the Plan Manager for a replanning process that will generate a new sequence of actions. This particular case would confirm that the child believes the proposed task is too dificult, and the robot will respond accordingly by implementing the Escape function strategy. However, this is just one possible scenario.
To assess our model’s efectiveness in educational settings, we have developed an experimental design comprising one-on-one child-robot interaction sessions, where the children will engage in a memory game task. The robot’s accuracy in detecting behavior functions and the appropriateness of employed strategies will be retrospectively evaluated by educational experts.
The experimental evaluation of the framework described in Sections 3 and 4, is conducted in a controlled laboratory setting focused on a memory game task (Fig.2b). The setup involves a oneon-one interaction between one of the authors and the humanoid robot Pepper. The experiment is repeated multiple times with the same participant to test various scenarios of replanning within the framework. The objective is to observe the triggering conditions for each scenario and evaluate whether the framework can accurately assert the relevant grounded predicates at the appropriate moments. To this end, key metrics being assessed include: • Replan Time: Measures the time from the introduction of a disruption to the completion of the replanning process. • Cloud Response Time: Evaluates the time taken for the robot to send a request, process cloud information, and receive a response.
Concerning Replan time, cloud resources were compared with local planning by sending the same PDDL problem to both a cloud-based Planner Engine and a local one. Planning times for the cloud range from 0.09 to 0.13 seconds, while local times range from 0.08 to 0.12 seconds. This indicates that cloud resources do not reduce planning time, likely due to the simplicity of the generated plans.
Regarding the Cloud Response Time, the graph in Figure 2a shows both the time required to retrieve a response sentence used to reply to the user (blue) and the time taken to retrieve grounded predicates (red). (a) Cloud Response Time vs. Number of Characters in the User’s Sentence. (b) Detecting Escape and Gain a Tangible during a robot-assisted memory card game.
The Figure 2a illustrates a diference in character processing between the two tasks. Our
implementation involves more complex instructions sent to OpenAI, incorporating additional context and
user-specific parameters. This leads to varying prompt lengths, impacting the average cloud response
time: approximately 1 second for sentence retrieval and 0.75 seconds for predicate retrieval. Both times
are significantly below the 2-second threshold deemed acceptable for human conversation [
Our study introduces a framework for enhancing child-robot interactions using behavior analysis principles. The system architecture includes an innovative framework for online replanning to adapt dynamically to user behavior. However, it is important to acknowledge several limitations. In its current initial version, our model does not account for children’s acceptance of the robot, their motivation, or attention during interactions. Addressing these factors is crucial for conducting more precise functional analyses. Future research should focus on optimizing our model by incorporating these elements to improve its efectiveness and acceptance.
This work has been co-funded by Ministero dell’Università e della Ricerca, Piano Nazionale Complementare (PNC), FIT4MEDROB project, Mission 3.