Assembly Sequence Planning via Spatio-Temporal State-Action Scoring

Abstract

We study assembly sequence planning on typed CAD joint graphs. Planning is formulated as a constrained longhorizon decision process. Each step selects a feasible part from the current frontier, and the process cost includes tool switching between consecutive attachments. We propose a constraintconsistent planner that separates feasibility from ranking. Frontier-based candidate generation enforces executability, while a spatio-temporal state-action scorer ranks feasible candidates. The scorer uses bilinear interactions between the current subassembly state and each frontier candidate, so it can model the relation between the last tool type and the next connection type. The learned scores are used as heuristics for greedy decoding and bounded beam decoding. To avoid manual sequence labels, we build a deterministic pseudo-expert from graph constraints and tool-switch cost. It generates reproducible demonstrations for step-wise imitation learning. Experiments on tasks built from Fusion 360 Assembly-Joint data show higher completion and fewer tool switches than structural heuristics and a linear scorer. The method remains stable under changes in cost weight and constraint noise.