BrickCraft: Visuomotor Skill Composition with Situated Manual Guidance for Long-Horizon Interlocking Brick Assembly

This paper presents BrickCraft, a compositional framework for long-horizon and generalizable interlocking brick assembly. It models the assembly process using a relative formulation, decomposing complex tasks into reusable primitive skills. Situated manuals provide explicit spatial guidance for learned visuomotor skills, and a modular execution pipeline enables the completion of long-horizon tasks. Experimental results show that BrickCraft effectively composes skills from limited demonstrations for diverse assembly tasks.