Learning to Scan: Solving 3-D Coverage Path Planning as a Traveling Salesman Problem via Reinforcement Learning

초록

Efficient automation of 3D scanning for large-scale structures demands globally optimized scan paths that satisfy sensor constraints. This paper proposes a TSP-Solving Transformer RL Planner (TS-TRP), a unified framework that reformulates the 3D surface coverage path planning (CPP) problem as a traveling salesman problem (TSP) and solves it via reinforcement learning (RL). The proposed framework first discretizes a complex 3D surface into uniformly-sized triangular mesh patches matched to the sensor specifications, thereby generating measurement nodes. It then employs Plücker coordinates from line geometry to compute the inter-node travel cost dictated by the local surface curvature in closed-form analytical expressions. A Transformer-based policy network is trained with the REINFORCE algorithm to learn a globally optimized scan ordering, which is subsequently refined through a 2-Opt local search in a hybrid optimization strategy. In real-world robotic experiments on an aircraft component (Leading Edge) with ten trials per method, TS-TRPLG, the variant augmented with line geometry waypoints, improved the Root Mean Square (RMS) registration accuracy by approximately 50% compared with expert manual scanning (3.98 mm →1.97 mm; Welch's t -test, p=0.003 ), and achieved significantly lower RMS than classical tour orderings combined with the same waypoint module (14-16% below 2-OptLG and ChristofidesLG, p< 0.001) together with 18-20% shorter scan times. In scalability experiments across six geometrically diverse surfaces, TS-TRP achieved path length reductions of up to 2.8%, and the application of line geometry waypoints reduced the mean midpoint standoff error (Mean MSE) by up to 96%. Ablation studies further demonstrated that the global tour ordering produced by TS-TRP contributes to scan quality improvement independently of the waypoint effect, thus confirming that the proposed framework simultaneously enhances both path efficiency and measurement quality for automated scanning of large-scale structures. © 2026 The Authors.

키워드