Prediction-Based Path Planning for Safe and Efficient Human–Robot Collaboration in Construction via Deep Reinforcement Learning

TLDR

Robotics is increasingly used in construction to assist workers with repetitive, physically demanding, and dangerous tasks, but in dynamic, complex workspaces robots must navigate efficiently without colliding with moving workers. This study proposes a deep reinforcement learning–based robot path‑planning method that incorporates predicted construction worker movements to achieve safe and efficient human–robot collaboration. The method first employs an uncertainty‑aware LSTM to predict worker trajectories and uncertainties, then embeds these predictions into a DRL framework’s state space and reward computation, and is validated in simulations derived from real construction scenarios. The prediction‑based DRL planner achieved a 100 % success rate over 10,000 episodes and reduced collision rates with moving workers by 23 % compared with conventional DRL, demonstrating proactive planning that improves safety and efficiency.

Abstract

Robotics has attracted broad attention as an emerging technology in construction to help workers with repetitive, physically demanding, and dangerous tasks, thus improving productivity and safety. Under the new era of human–robot coexistence and collaboration in dynamic and complex workspaces, it is critical for robots to navigate to the targets efficiently without colliding with moving workers. This study proposes a new deep reinforcement learning (DRL)–based robot path planning method that integrates the predicted movements of construction workers to achieve safe and efficient human–robot collaboration in construction. First, an uncertainty-aware long short-term memory network is developed to predict the movements of construction workers and associated uncertainties. Second, a DRL framework is formulated, where predicted movements of construction workers are innovatively integrated into the state space and the computation of the reward function. By incorporating predicted trajectories in addition to current locations, the proposed method enables proactive planning such that the robot could better adapt to human movements, thus ensuring both safety and efficiency. The proposed method was demonstrated and evaluated using simulations generated based on real construction scenarios. The results show that prediction-based DRL path planning achieved a 100% success rate (with a total of 10,000 episodes) for robots to achieve the destination along the near-shortest path. Furthermore, it reduced the collision rate with moving workers by 23% compared with the conventional DRL method, which does not consider predicted information.