Assist-as-needed robotic trainer based on reinforcement learning and its application to dart-throwing

Area of Science:

• Robotics
• Motor Learning
• Human-Computer Interaction

Background:

• Robotic assistance in motor learning often uses fixed trajectories, potentially suboptimal for individual users.
• The guidance hypothesis highlights the risk of over-reliance on external robotic feedback, hindering internal skill acquisition.
• Existing adaptive systems often require complex user physical models, limiting practical application.

Purpose of the Study:

• To propose a novel user-adaptive robotic trainer for motor skill learning.
• To develop a system that avoids predetermined trajectories and complex user models.
• To leverage model-free reinforcement learning for adaptive robotic assistance.

Main Methods:

• Developed a user-adaptive robotic trainer framework.
• Implemented model-free reinforcement learning for adaptive control.
• Utilized dart-throwing as a quantifiable motor-learning task.

Main Results:

• Demonstrated the feasibility of the proposed framework through training experiments with novices.
• Showcased the system's ability to adapt assistance based on user performance.
• Achieved success in maximizing task scores while minimizing robotic assistance.

Conclusions:

• The proposed model-free reinforcement learning approach offers a viable method for user-adaptive robotic trainers.
• This framework effectively supports motor skill acquisition without requiring user-specific physical models or trajectories.
• The system shows promise for enhancing physical therapy and motor rehabilitation applications.