Abstract
Combinatorial assembly uses standardized unit primitives to build objectsthat satisfy user specifications. Lego is a widely used platform forcombinatorial assembly, in which people use unit primitives (ie Lego bricks) tobuild highly customizable 3D objects. This paper studies sequence planning forphysical combinatorial assembly using Lego. Given the shape of the desiredobject, we want to find a sequence of actions for placing Lego bricks to buildthe target object. In particular, we aim to ensure the planned assemblysequence is physically executable. However, assembly sequence planning (ASP)for combinatorial assembly is particularly challenging due to its combinatorialnature, ie the vast number of possible combinations and complex constraints. Toaddress the challenges, we employ deep reinforcement learning to learn aconstruction policy for placing unit primitives sequentially to build thedesired object. Specifically, we design an online physics-aware action maskthat efficiently filters out invalid actions and guides policy learning. In theend, we demonstrate that the proposed method successfully plans physicallyvalid assembly sequences for constructing different Lego structures. Thegenerated construction plan can be executed in real.