[Paper Review] Personalized Behaviour Models: A Survey Focusing on Autism Therapy Applications

Michał Stolarz, Alex Mitrevski+2 more|arXiv (Cornell University)|May 18, 2022

Autism Spectrum Disorder Research4 citations

TL;DR

This survey proposes a comprehensive analysis of personalized behavior models in robot-assisted therapy (RAT) for children with Autism Spectrum Disorder (ASD), focusing on interactive machine learning (IML) techniques—learning from guidance and feedback—to enable robots to autonomously adapt to individual children's behaviors. The key contribution is a comparative evaluation of personalization approaches across social behavior, task difficulty, and engagement, identifying learning from feedback and hybrid learning as most feasible for real-world autism therapy applications with faster convergence and reduced reliance on expert supervision.

ABSTRACT

Children with Autism Spectrum Disorder find robots easier to communicate with than humans. Thus, robots have been introduced in autism therapies. However, due to the environmental complexity, the used robots often have to be controlled manually. This is a significant drawback of such systems and it is required to make them more autonomous. In particular, the robot should interpret the child's state and continuously adapt its actions according to the behaviour of the child under therapy. This survey elaborates on different forms of personalized robot behaviour models. Various approaches from the field of Human-Robot Interaction, as well as Child-Robot Interaction, are discussed. The aim is to compare them in terms of their deficits, feasibility in real scenarios, and potential usability for autism-specific Robot-Assisted Therapy. The general challenge for algorithms based on which the robot learns proper interaction strategies during therapeutic games is to increase the robot's autonomy, thereby providing a basis for a robot's decision-making.

Motivation & Objective

• To analyze and compare personalized behavior models in Human-Robot Interaction (HRI) for autism therapy applications.
• To evaluate the feasibility and usability of different personalization techniques—especially learning from feedback and guidance—in real-life Robot-Assisted Therapy (RAT) scenarios for children with ASD.
• To identify key challenges in deploying autonomous, adaptive robots in long-term therapy, including reliance on human supervisors, slow policy convergence, and generalization across users.
• To guide future development of autonomous robots that can personalize therapy content and respond to child disengagement or demotivation in real time.
• To propose future work focused on fast-converging, pretrained behavior models using active learning and hybrid learning (feedback + guidance) for improved therapy outcomes.

Proposed method

• The survey analyzes 18 representative studies on personalized robot behavior models in HRI, categorized by personalization aspect (e.g., social behavior, task difficulty) and learning method (learning from feedback or guidance).
• It evaluates each approach based on learning algorithm type (e.g., Q-learning, Bayesian networks, nearest neighbors, inverse reinforcement learning), features used (e.g., engagement, game progress, gaze, speech), and whether long-term operation (>1 week) was demonstrated.
• The authors compare techniques in terms of autonomy, robustness to supervisor errors, convergence speed, and generalization across users, using a structured comparison table (Table I).
• The methodology includes identifying deficits in current approaches, such as lack of long-term testing, reliance on expert supervision, and slow learning, and proposes hybrid learning (feedback + guidance) as a solution.
• The survey draws on existing taxonomies from HRI literature (e.g., user model types: static vs. dynamic) to structure the analysis and assess applicability to autism therapy.
• Future work is proposed to develop a robot that uses active learning and policy pretraining to enable fast, personalized adaptation to individual child skills and emotional states.

Experimental results

Research questions

• RQ1 Which personalization techniques in robot-assisted therapy for children with ASD are most feasible for real-world deployment?
• RQ2 How do learning-from-feedback and learning-from-guidance approaches compare in terms of robustness, convergence speed, and adaptability in long-term therapy sessions?
• RQ3 What are the key challenges in deploying autonomous, personalized behavior models in autism therapy robots, particularly regarding supervision dependency and generalization across users?
• RQ4 Can hybrid learning approaches (combining feedback and guidance) improve the reliability and speed of policy learning in therapy robots?
• RQ5 How can personalized behavior models be designed to detect and respond to child disengagement or demotivation in real time?

Key findings

• Learning from feedback is more feasible than learning from guidance for long-term, real-world autism therapy, as it reduces dependency on human supervisors and allows autonomous exploration.
• Only 3 out of 18 surveyed approaches demonstrated long-term operation (>1 week), indicating a significant gap in real-world validation of personalized behavior models.
• Hybrid learning approaches (e.g., combining feedback and guidance) were found to improve robustness to supervisor errors and accelerate policy convergence, as shown in studies like [27] and [41].
• The most effective personalization dimensions in autism therapy are social behavior and task difficulty, with social behavior personalization being particularly impactful for maintaining engagement.
• Pretraining policies on simulated user interactions can significantly reduce the number of real interactions needed for convergence, addressing the challenge of slow learning in real therapy settings.
• Inverse reinforcement learning and model-based reinforcement learning show promise for capturing complex user preferences and emotional states, though they require careful feature engineering and are less tested in long-term scenarios.