Most automated assembly research focuses on "semantic" tasks—putting together a chair from a leg, a seat, and a backrest. In these scenarios, the machine knows what the final object should look like and what role each part plays. However, in archaeology, forensics, and disaster recovery, we face the "jigsaw puzzle from hell": reassembling objects shattered into hundreds of irregular, non-semantic shards.
The emerging research direction in Geometric Fracture Assembly moves away from top-down category knowledge. Instead, it treats reassembly as a bottom-up problem of finding a "lock and key" fit between jagged surfaces.
The Shift from Semantics to Pure Geometry
Traditional assembly models rely on semantic labels—knowing that a specific part is a "table leg" helps determine its orientation. But as Jiaxin Lu and Qixing Huang point out in Jigsaw, fractured shards lack these labels. A piece of a shattered Ming vase doesn't look like a "handle" or a "rim"; it looks like a generic, irregular lump.
To solve this, researchers like Silvia Sellán and Yun-Chun Chen introduced the Breaking Bad dataset. This work formalizes the problem: reassembly must prioritize local surface matching over global appearance. Because individual shards are often unrecognizable, the assembly logic must be extracted from the fine-grained geometry of the fractured edges themselves.
The Lock and Key Mechanism
The core intuition behind modern reassembly is identifying "breaking curves"—the jagged lines where a fracture occurred. Ali Alagrami and Marcello Pelillo demonstrate in Reassembling Broken Objects using Breaking Curves that by segmenting fragments based on these curves, algorithms can perform an exhaustive search for complementary geometric features.
This is often framed as a "mating" problem. Neural Shape Mating, developed by Yun-Chun Chen and colleagues, uses a self-supervised Transformer to "snap" two parts together. Instead of looking for identical patterns, the model learns to recognize how a convexity in one piece fits into a concavity in another. This allows the system to generalize to entirely novel objects it has never seen before, as it isn't looking for a "vase" shape, but rather a physical fit.
Scaling to Multi-Fragment Puzzles
While mating two pieces is a start, real-world problems involve dozens of fragments. This introduces a combinatorial explosion: every piece could potentially connect to every other piece.
To manage this complexity, Ruiyuan Zhang and Hao Dong proposed a Co-creation Space. This framework acts as an information bottleneck, allowing multiple "assembler" agents to share information about potential fits. By shifting from quadratic to linear complexity, they enabled the assembly of up to 20 fragments simultaneously.
Similarly, Wenting Cui and Shaoyi Du introduced PHFormer, which represents fragments as sets of local geometric "proxies." This hierarchical approach allows the model to reason about adjacency and global structure without the computational burden of matching every individual point in a point cloud.
Global Consistency from Local Chaos
Even without semantic labels, fragments often follow global geometric rules. For example, many archaeological artifacts are axially symmetric (like pots or bowls). Kilho Son and David B. Cooper showed in Axially Symmetric 3D Pots Configuration System that fragments can be aligned by identifying a shared axis of symmetry, even if the pieces are not physically adjacent.
This "geometric consensus" is a powerful tool. In Solving Small-Piece Jigsaw Puzzles by Growing Consensus, Kilho Son demonstrates that by forming stable loops and grids of pieces, an algorithm can self-correct errors. If three pieces fit together but the fourth doesn't close the loop, the system knows the previous matches were likely false positives.
Why This Works
The success of these methods suggests that local geometric compatibility is a robust proxy for global structure. By focusing on the "how" (the physical fit of fracture surfaces) rather than the "what" (the category of the object), these models achieve three things:
- Generalization: They work on novel, irregular objects that don't fit into standard categories.
- Robustness: They can handle missing data and eroded edges by relying on sparse pictorial or geometric cues.
- Automation: They reduce the need for manual handling of delicate physical artifacts, as seen in the work of Xinda Liu and Guohua Geng on Cultural Heritage Protection.