Towards a Human-Sensitive Controller: Learning Human Specificities in Ergonomics and Physical Constraints

Work-Related Musculoskeletal Disorders remain a major challenge in industrial settings, accounting for over half of occupational diseases in Europe and imposing economic and social burdens. These disorders often cause chronic pain and reduced work capacity, highlighting the need for inclusive workstations that adapt to workers' physical constraints. This paper explores Reinforcement Learning (RL) for developing a personalized control strategy for a collaborative robot (cobot). Q-Learning enabled a cobot to optimize human ergonomic posture while considering physical constraints. This model-free approach allows the cobot to learn optimal actions through interaction with the environment, maximizing a reward function designed to minimize ergonomic and pain risk levels. To evaluate discretization effects, two state space levels (10 cm and 6.25 cm) were tested. Models were initially trained in simulation and fine-tuned in real-world settings, Results underscore the importance of fine-tuning policies to bridge the sim-to-real gap. Fine-tuned policies eliminated pain risk and ensured safe ergonomic postures. Performance was evaluated using reward per episode, ergonomic and pain risk levels, and steps per episode, demonstrating RL-driven cobots' potential to enhance worker health and inclusion.