A hybrid deep learning and meta-reinforcement learning architecture for adaptive AI-assisted fine motor control training in older adults

Against the backdrop of global population aging, declining fine motor control in older adults significantly impairs their ability to operate digital devices, directly limiting their participation in digital learning activities and smart elderly care services. Traditional motor training methods lack personalization and real-time adaptability, while imposing high service costs that strain elderly care management systems. This study proposes a hybrid intelligent architecture integrating state representation learning and meta-reinforcement learning to address these challenges. The architecture extracts low-dimensional state embedding vectors from joystick interaction data through a deep encoding network, accurately quantifying participants’ motor control capabilities and task progress. An adaptive decision-making module based on meta-reinforcement learning enables rapid adaptation to individual differences among older adults, achieving dynamic balance between intervention effectiveness and autonomy cultivation. A multi-dimensional reward function integrates task completion, user autonomy, intervention intensity, and motor skill progress into a unified optimization objective. Experimental results demonstrate that the system significantly improves participants’ performance in trajectory-tracking tasks and motor skill acquisition efficiency. User autonomy ratio increases progressively throughout training while AI intervention gradually decreases, validating the hybrid architecture’s effectiveness in optimizing human-AI collaborative training. This research provides technological pathways for promoting active aging, bridging the digital divide, and enhancing operational efficiency in elderly care service management.