RL-Lock: Reinforcement Learning for Generating Interlocking Assemblies

Generating interlocking assemblies—structures held together solely by their geometry without adhesives or fasteners—is a classic challenge in computational design. Traditionally, this is treated as a shape decomposition problem, where a 3D object is partitioned into pieces. Existing methods rely on manually designed search heuristics to navigate the massive combinatorial space of possible voxel assignments. This paper asks whether a reinforcement learning (RL) agent can learn to generate these assemblies from experience, potentially surpassing the performance of heuristic-based approaches in complex, high-piece-count scenarios.

The authors frame the generation of interlocking assemblies as a sequential decision-making process. They model this as a Finite Markov Decision Process (MDP), where an agent iteratively assigns voxels to pieces. To handle the long-horizon nature of this task, the researchers introduce two key innovations:

1. Structured Action Chunking: Instead of making decisions at the individual voxel level, the framework groups sequences of voxel assignments into high-level "action chunks" that represent the construction of a single piece. This significantly reduces the effective decision horizon.
2. MCTS-Guided Learning: The framework employs a policy-value neural network trained via Monte Carlo Tree Search (MCTS). The network learns to predict the probability of successful action chunks and the expected value of a given state, allowing the MCTS to prioritize promising geometric configurations.

RL-Lock demonstrates that an RL-based approach can effectively generate valid interlocking assemblies. By learning from previous search episodes, the model avoids the limitations of rigid, human-designed heuristics. The experiments show that RL-Lock successfully generates complex assemblies—such as 9, 16, and 25-piece cubes—in challenging cases where traditional heuristic-based search methods either fail to find a solution or require prohibitive amounts of computation time.