— This paper presents a new method for planning motions of multi-arm systems in constrained workspaces, for which state-of-the-art planners behave poorly. The method is based on the decomposition of the system into parts. Compact roadmaps are first computed for each part, and then, a super-graph is constructed by the composition of elementary roadmaps. Results presented for a three-arm system and a model of the complex DLR’s Justin robot show a significant performance gain of such a two-stage roadmap construction method with respect to single-stage methods applied to the whole system.