Learning-based personalisation of robot behaviour for robot-assisted therapy.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Robot-assisted therapy often requires significant therapist control (e.g., Wizard-of-Oz protocol), limiting therapist focus on patient interaction.
• Increasing robot autonomy can alleviate therapist burden and potentially improve therapeutic session efficiency.

Purpose of the Study:

• To develop a learning-based behavior model for increased robot autonomy in therapy.
• To investigate reinforcement learning techniques and compare reward functions for user engagement and performance.
• To analyze strategies for more tractable learning, including user modeling, policy transfer, and expert feedback.

Main Methods:

• Utilized reinforcement learning to train robot behavior models.
• Compared various reward functions focusing on user engagement and task performance.
• Investigated policy transfer between user groups and learning from expert feedback.
• Evaluated learned models in a small-scale feasibility study using a sequence learning game.

Main Results:

• Policy transfer demonstrated a significant acceleration of the policy learning process.
• The choice of reward function critically influences the robot's action selection.
• Initial feasibility study results suggest learning from guidance may yield optimal policies for user engagement and game performance.

Conclusions:

• A learning-based approach can enhance robot autonomy in therapeutic settings.
• Policy transfer is an effective strategy for speeding up robot behavior learning.
• Further large-scale studies are required to validate the effectiveness of guidance-based learning for improving user outcomes.