New AI Teaches Robots to Assemble Puzzles Like Tangrams
This could lead to robots that build furniture, electronics, and more—without frustration.
Researchers present an object-centric reinforcement learning framework that treats each part as an agent and learns a shared policy for continuous assembly control. A graph neural network represents the pieces, their target poses, and their spatial relationships, and predicts translational and angular velocities for each component.
Per-agent proximal policy optimization with a centralized graph critic learns motion and temporal coordination within a single feedback policy. They study this formulation through planar Tangram assembly, where pieces must slide and rotate into a given configuration while remaining on the assembly surface, using procedurally generated target configurations and randomized initial arrangements.
Experiments isolate the roles of relational structure and multi-agent optimization and compare learned control with sampling-based planning. They demonstrate concurrent assembly in simulation and transfer the policy to sequential execution on a real robot, because the object-centric formulation decouples learned assembly coordination from low-level physical control.
- The AI treats each puzzle piece as an independent agent and learns to coordinate their movements.
- It uses a graph neural network to understand spatial relationships and avoid collisions.
- The system was tested on Tangram puzzles and transferred to a real robot, showing potential for flexible assembly.
Why It Matters
This could lead to robots that assemble products faster and adapt to new tasks, saving time and money in manufacturing.