Neural dynamic assembly sequence planning

The automatic generation of feasible assembly sequences from CAD-data is for several reasons a challenging task. One reason being that with increasing number of parts of an assembly group, the number of possible sequences grows exponentially, making an exhaustive search non-practical. We face this combinatorial problem by using Reinforcement Learning (Deep-Q-Learning) to approximate the cost-function of the assembly with an artificial neural network (ANN) and guide the search for an asymptotically optimal solution of this combinatorial problem. Assembly costs are calculated with a collision-based assembly-by-disassembly approach. The derived method is tested on assemblies of different sizes and types. The presented method provides collision-free assembly sequences very fast, due to its depth-first character and solves small and medium tasks reliably.