Optimizing robotic connection task sequencing in off-site structural assembly: an obstacle-aware deep reinforcement learning method

Robotic connection operations, such as nailing and screw fixing, are increasingly feasible in off-site structural assembly. However, efficient task sequencing remains challenging because connection tasks are numerous, spatially distributed, and constrained by tool capacity and physical obstacles. This study proposes an obstacle-aware deep reinforcement learning (DRL) framework for robotic connection task sequencing under tool-capacity constraints. The sequencing problem is formulated as an adaptation of a classical capacity-constrained routing problem, and shortest-path distance matrices are integrated into the DRL framework to represent obstacle-aware travel costs. These matrices are precomputed before training, decoupling CPU-based shortest-path calculations from GPU-based DRL training and thereby making the training process computationally feasible. The framework is implemented using two attention-based DRL algorithms, Attention Model (AM) and Policy Optimization with Multiple Optima (POMO), and evaluated against the exact solver Gurobi and three heuristic methods. Across single-frame and multi-frame off-site structural assembly scenarios, both DRL models outperformed all heuristic methods and required less than 0.1 s per instance for inference. The best-performing DRL algorithm achieved average solution gaps of 0.84 % and 2.79 %, respectively, relative to the best-found Gurobi solutions in two scenarios. Overall, the results demonstrate the effectiveness of the proposed framework for fast and high-quality task sequence optimization, representing the first application of an integrated attention-based DRL and shortest-path planning framework to robotic connection operations in off-site structural assembly.