Why Assembly Planning Is a Frontier Problem in Mechanical Engineering

The Missing Piece in Physical AI

AI has mastered language and vision, but it still cannot plan how things fit, move, and come together. In other words, AI does not have mechanical intuition. Until this bottleneck is solved, every attempt to automate the physical world will fail. The race is on—NVIDIA and others are training foundation models for robotics, China is compressing the design‑to‑manufacturing loop with vertically integrated, state‑backed automation, and spatial intelligence is emerging as AI's next frontier. Advanced manufacturing is the battleground where these forces will converge to shape geopolitical influence.

At C‑Infinity we are building AI with mechanical intuition. The most acute need is in advanced manufacturing, where the hardest mile in product realization lies between design intent and shop‑floor execution -- the challenge is to convert product structure into process structure. We can frame this problem in terms of how the view of a product changes as it progresses from design to manufacturing.

• As‑designed view: the design engineer's model (in CAD) capturing nominal (perfect-form) geometry and grouping subassemblies by function. This is the design intent.
• As-planned view: the manufacturing engineer's model capturing materials, fixtures, tools, and build tasks, capturing the steps required to transform raw material into a finished product.
• As‑manufactured view: The result of translating the tasks in the virtual build into a physical build.

Bridging the gulf between a product's as‑designed and as‑manufactured views remains one of mechanical engineering's toughest challenges. Translating one into the other requires production planning, where spatial planning meets design intent and manufacturing constraints—resulting in a virtual build that must succeed before a single bolt is tightened on the factory floor. The virtual build maps the entire sequence of production steps, from raw material to finished product, representing the as-planned view.

Assembly Planning: Complexity That Needs Taming

Unlike machined parts—where CAM systems automate tool paths, though engineers still plan setups and select tools—there is no equivalent CAM system for assemblies. Assembly planning, which determines the sequence of part and subsystem installations, remains largely manual or ad hoc. The closest digital aids are exploded views or basic motion planning tools, but these still require the user to define the installation order. Ultimately, a human or AI planner must:

1. Divide the product into intermediate sub‑assemblies
2. Determine a feasible insertion order guaranteeing collision‑free, stable motions for every component and intermediate sub‑assembly.
3. Design or select fixtures that keep each transient state rigid and safe.

What changes between parts and assemblies is not merely scale; it is a step‑function increase in combinatorial, geometric, and physical complexity. This higher‑order problem of assembly planning is where progress has stalled.

Every modern product is an assembly; therefore every manufacturer endures this gap. Without a reliable virtual build, teams are forced to iterate with physical prototypes—machined one‑offs, 3‑D‑printed fixtures, and trial assembly stations—to expose clearance clashes, stability issues, and human‑factor surprises that digital tools still miss. These prototypes burn calendar, material, and cash, yet they remain the industry's only insurance policy against catastrophic late‑stage failures.

Why Assembly Planning Defies Classic Approaches

Since the earliest days of robot motion planning starting from the classic work of Lozano-Perez that laid the foundation for the field, researchers have tried to extend spatial planning and reasoning to manufacturing challenges—automated grasping, intelligent fixturing, and, most critically, assembly planning. Work by Bourjault, de Mello and Sanderson, Whitney, Latombe, Tian et al, and many others have made notable progress, yet commercial PLM/MES platforms still delegate planning to humans. The core reasons are:

1. Combinatorial Explosion. A 20‑part assembly has 20! (> 2×10¹⁸) theoretical build orders—yet only a microscopically small subset obeys geometric clearance, stability, resource, and ergonomic constraints. Finding a good plan is like finding a needle in an astronomical haystack.
2. Rich, Non‑Convex Geometry. Real parts have complex, non-convex shapes. Classic analytical research in fixturing, robot motion planning, grasping, and assembly planning all have similar mathematical roots, but implementations and formulations represent parts as juxtapositions of convex polyhedra.
3. Path & Ergonomic Feasibility. Whether the assembly is manufactured by robots or human hands, there must be a collision free path to move the part into the assembly achievable by the robot or human hand. It is possible to use motion planning and physical (numerical) simulations to solve the problem for smaller assemblies but the approach will not scale to assemblies beyond tens of parts.
4. Variant Proliferation. Most advanced manufacturing products are configurable, and the number of configurations are in the millions or billions per enterprise. In practice, OEMs define a configurable product as a 200% EBOM, including all variants of subassembly configurations in a single CAD design. Mapping this 200% EBOM to a 200% MBOM is unsolved.
5. Ambiguous Design Intent. Design intent can be hard to identify. Every CAD file we have ever encountered has multiple errors such as duplicate parts, mismatched threads, intended interferences (when modeling threads) and unintended interferences, etc. The assembly planner must determine what the designer intended before developing a manufacturing plan.
6. Flexible & Press‑Fit Components. Many parts are flexible. A part that snaps into an existing assembly requires its model to be nonrigid. Deformable gaskets are often used between engine parts. Material properties are typically not included in CAD files.
7. Physics & Stability. Current design tools are primarily graphics editors, allowing designers to create impossible designs. Design software does not understand mechanical engineering across disciplines. Geometry matters. The assembly plan must be precise. The tiniest error, gap, instability, interference will make the assembly unmanufacturable on the factory floor. Physics matters. Putting an assembly together that falls over or breaks during construction due to intermediate overstress is a nonstarter. Running a finite element analysis or any other physical simulation (e.g. for contact modeling) will significantly stress the ability to explore any meaningful search space.
8. Fixture Design. Fixtures need to be designed. All designs need fixtures which the designer rarely considers. At the very least we need to design plans with stable configurations.
9. Subassembly Re‑factoring. Designs always have subassemblies. But for manufacturing subassemblies often must be adjusted to make efficient manufacturing possible.

What Changed — Three Catalysts

1. Commodity Parallel Compute. GPUs bring teraflops of collision and physics simulation. We use GPUs to rapidly map and explore the feasible solution space for assembly plans.
2. Neuro‑Symbolic AI. ML extracts intent from CAD; symbolic solvers ensure deterministic and collision free planning. We also use surrogate physics models in the planner to guide the search.
3. Graph‑Native Data Models. We transform CAD assemblies into searchable graphs, enabling millisecond queries across feasible states. This allows users to interactively edit plans and replan or validate plans in seconds.

Together these catalysts render industrial‑scale assembly planning tractable. We have demonstrated AutoAssembler's ability to plan on realistic industrial assemblies with hundreds to thousands of parts in minutes.

AutoAssembler — The Missing Compiler

Assembly planning is where bits meet atoms. Solving it demands the combined rigor of computational geometry, AI, and domain expertise. Transforming the as-designed view to the as-planned view is a 'compilation' problem transforming a higher level language (design models) into lower level execution instructions. As with compilers, this requires reorganization of the source code (subassembly hierarchies) and optimizing the instruction set for the execution architecture (the factory floor). AutoAssembler is the missing compiler needed to build the intelligence layer between design and manufacturing at the product assembly scale. By solving this frontier problem, we provide US manufacturers with enterprise-grade tools to rapidly transform product designs into production-ready assemblies with detailed instructions. Contact us to learn more, partner, and build the future of advanced manufacturing.