Enhancing Collision Avoidance in Construction Robot Manipulation Using Trust Region Conditional Value at Risk

Emerging construction robots enhance workplace safety but also introduce new physical collision risk for workers. While learning-based robot control has demonstrated promising performance in tackling complex tasks, limited research has focused on ensuring the safety of neural network policies. This study addresses this gap by employing a safe deep reinforcement learning approach, incorporating conditional value-at-risk constraints instead of expectation-based constraints to prioritize safety in human–robot collaboration within construction environments. This method outperforms previous reinforcement learning algorithms that utilize collision penalty rewards for collision avoidance. Static and dynamic construction environments have been used to evaluate collision avoidance, and results show that the used method can significantly reduce the collision rate, thus improving construction safety.