Assembling objects from multiple parts is tricky when pieces must move in tight spaces and risk blocking each other. This coordination challenge appears in tasks like furniture assembly or packing.
The authors treat each puzzle piece as an independent agent that learns to move cooperatively. A graph neural network lets pieces share information about their positions and targets, and a shared policy trained with multi-agent decides each piece's velocity.
The learned policy successfully assembles Tangram puzzles in simulation and on a real robot. It coordinates concurrent piece motions while avoiding collisions, demonstrating that multi-agent learning can solve geometric assembly without explicit planning.