Reinforcement Learning Based Decision Support for Adaptive Pacing Control of Hand Activity Level

Abstract

This paper presents a reinforcement learning based decision support system for adaptive pacing control in manual material handling tasks. Hand Activity Level optimization is formulated as a finite horizon Markov Decision Process, where actions select pacing intensities or micro breaks under stochastic object loads. Q-learning and Deep Q-learning are evaluated in simulation using productivity and cumulative ergonomic exposure metrics aligned with HAL-TLV guidelines. Results show that Deep Q-learning achieves higher productivity, but Q-learning has lower cumulative risk exposure.