Keeping Assemblies in One Piece: A Connectivity Barrier for Collision-Free Multi-Part Placement

When a model places the parts of a 3D assembly, such as the brackets, shafts and housings of a robot arm, it tends to fail in two opposite ways. Parts pass through each other, or they float apart with no physical support. The usual fix for the first failure causes the second. A penalty on overlap pushes parts away from each other, and once a part has drifted out of contact nothing pulls it back. The OmniCAD benchmark measures how bad this is. Across 25,000 industrial assemblies, frontier vision-language models produce interpenetrating parts in more than half of their outputs and reach at most 46.1% PairF1 when recovering which parts mate with which.

This note proposes treating "floating" as a connectivity failure, in the same sense that modular-robotics researchers mean when a module becomes stranded during reconfiguration. In that literature, connectivity is enforced on every intermediate configuration, not only checked at the end, because a module that has detached cannot rejoin the structure. The proposal borrows that discipline for the optimizer that refines an assembly. It builds a smooth "contact graph" whose edge weights shrink as the clearance between two parts grows. It then adds a barrier that keeps the graph's algebraic connectivity, a single number measuring how well the graph holds together, above a threshold at every optimization step. A compliant penetration penalty does the opposite job, removing overlap without demanding exact zero clearance. The claim to test is narrow: when the list of which parts should touch is unreliable, as it is for every current predictor, this global connectivity barrier prevents floating better than attraction terms placed on individual part pairs, and it does so without bringing interpenetration back.

Core Claim and Mechanism

Each part $i$ of an assembly with $K$ parts has a rigid pose $T_i$. For each pair of parts, $d_{ij}$ is their signed clearance, computed from part signed distance fields (distance-to-surface functions, negative inside a part). The clearance is positive when the parts are apart and negative when they overlap. The contact weight is $w_{ij} = \exp(-\max(d_{ij},0)^2/\sigma^2)$, where $\sigma$ sets how quickly contact credit decays with distance. These weights define a graph Laplacian $L(w)$, and its second-smallest eigenvalue $\lambda_2$ is positive exactly when the weighted contact graph is connected. The refinement objective is

$$ E(T) = \sum_{(i,j)\in \hat M} c_{ij}\, E^{\text{mate}}_{ij}(T) + \alpha \sum_{i<j} \max(-d_{ij} - \delta, 0)^2 - \beta \log\big(\lambda_2(L(w(T))) - \epsilon\big), $$

where the terms are defined as follows. $\hat M$ is the set of predicted mates, and $c_{ij}$ is the confidence of each mate. $E^{\text{mate}}_{ij}$ is the geometric residual of a mate type, for example a concentric or coincident constraint. $\delta$ is a penetration allowance. $\alpha$ and $\beta$ weight the penalty terms, and $\epsilon$ is the minimum connectivity the barrier enforces.

The gradient of $\lambda_2$ with respect to an edge weight is $(v_i - v_j)^2$, where $v$ is the Fiedler vector, the eigenvector associated with $\lambda_2$. The barrier therefore strengthens exactly the edges that bridge the weakest cut in the assembly, whichever parts those happen to be.

Three prior lines of work supply the pieces.

The barrier comes from Rigidity Maintenance Control for Multi-Robot Systems. That work keeps a spectral rigidity eigenvalue above a minimum during swarm motion using a potential that diverges near the threshold. Its open problem is the cost of tracking the global eigenvector across hundreds of distributed robots. That cost disappears here, because an assembly averages about 12 parts and the eigenproblem is solved centrally.

The penetration allowance $\delta$ comes from ContactOpt. ContactOpt found that permitting up to 2 mm of interpenetration gives stable gradients and more realistic grasps than strict zero-penetration constraints. For rigid metal parts, $\delta$ will be much smaller, and it will be tuned.

The paired attract-and-repel structure comes from CPF. CPF uses attractive springs on predicted contact pairs and repulsive springs along surface normals. The proposal differs from it in one respect: support is required over the whole graph rather than per pair, so a wrong or missing pair prediction does not strand a part.

The reason to enforce connectivity along the whole trajectory, and not only at the final configuration, comes from the reconfiguration literature. In the connectivity-preserving grid transformations of On efficient connectivity-preserving transformations in a grid, every intermediate shape must stay connected because detached modules lose power and communication. An optimizer has an analogous problem. Once a part's clearance passes a few multiples of $\sigma$, its contact weight and the gradient of that weight both vanish, and the part is stranded.

Research Program

The work builds a refinement module that takes any initial assembly proposal and returns adjusted part poses. Initial proposals will come from the strongest model's one-shot OmniCAD outputs and, separately, from a simple learned pose regressor. The predicted mate set $\hat M$ will come from the same model, so the mates carry realistic errors.

Signed distance fields are precomputed per part from the B-Rep meshes. Overlap is estimated by dense surface sampling, which addresses the vertex-resolution weakness noted in the research file: a sharp vertex can slip between sparse face samples without triggering a penalty. Optimization runs on the 6-DoF poses with a continuation schedule. The barrier is active from the first step, and the threshold $\epsilon$ is tightened gradually.

The analysis component asks when the global barrier is necessary rather than merely convenient. The main instrument is a controlled sweep over mate-graph quality. It starts from ground-truth mates and progressively drops true mates and injects false ones, until the mate set matches the predictor's real 46.1% PairF1.

The proposal also states where it is not expected to help. The barrier does nothing about deep initial penetrations, where repulsive gradients can push a part out through the wrong face. Those cases will be reported separately, not absorbed into averages.

Baselines

The do-nothing reference is the unrefined one-shot output of the best OmniCAD model. The paper reports 27.9% pose accuracy within 10 mm, more than 50% of assemblies with interpenetration, and 46.1% mate PairF1 for this output. Every refinement method starts from these exact proposals.

The strongest existing method that targets interpenetration on this benchmark is OmniCAD's agentic loop, in which the model sees renders and collision reports and revises coordinates. The paper reports only that this loop helps coarse alignment, gives little gain in fine-grained alignment, and raises runtime from about 50 seconds to nearly 5 minutes per sample. The file contains no post-loop interpenetration rate, so there is no reported result found on Lacuna for that metric. The loop will be rerun under its published protocol, and the comparison will report both quality and wall-clock time.

The strongest optimization competitor is a pairwise contact-spring refiner in the style of CPF. It uses the same predicted mates, the same signed distance fields, the same compliant penetration term and the same optimizer, and adds attractive springs on every predicted mate. It receives an equal hyperparameter search budget on a held-out split. No reported result was found on Lacuna for any differentiable refiner on OmniCAD, so this baseline is new to the benchmark and will be tuned as carefully as the proposed method.

A repulsion-only refiner, with penetration penalty and mate terms but no attraction, isolates the floating failure described in Layout-your-3D.

Two ablations isolate the key ingredient. In the first, the same $\lambda_2$ term is applied only at the final iterate as a penalty rather than as a barrier active throughout. In the second, the spectral term is replaced by attraction along a minimum spanning tree of current clearances, which is cheap, non-smooth, and recomputed every step.

Beating these baselines means a lower floating rate than the pairwise-spring refiner at matched or lower interpenetration rate, with no loss in 10 mm pose accuracy, under predicted mates.

Why not just…?

Why not just put attractive springs on the predicted mate pairs? This is the real threat. With correct mates it may match the barrier, because pairwise attraction along a connected mate graph already implies global connectivity. The case for the barrier depends entirely on mate errors. A missing mate leaves a part with no attraction at all, and a false mate pulls a part to the wrong neighbor. The mate-quality sweep settles the question. If pairwise springs match the barrier at the predictor's realistic error level, the proposal's claim fails, and that result will be reported.

Why not just settle the assembly in a physics engine? Simulators resolve overlap, but they push overlapping parts apart with impulses. The CPF evaluation in Bullet found that baselines dropped or launched objects. Without fixtures, a settled assembly scatters or collapses under gravity, so penetration is traded directly for disconnection. Bullet settling will be run as an additional reference on the same proposals, with floating measured afterward.

Why not just attract each part to its nearest neighbor, or along a spanning tree? Nearest-neighbor attraction can leave two internally connected clusters that never join. Spanning-tree attraction does fix global connectivity, but its tree changes discontinuously as clearances cross each other, which may cause oscillation. It is included as an ablation. If it matches the spectral barrier, the simpler rule should be preferred, and the note will say so.

Why not just iterate with an agent? OmniCAD already reports little fine-grained gain from the agentic loop at roughly six times the runtime, and the rerun will check whether that holds on this proposal's metrics.

Proposed Experiments

All experiments use OmniCAD's assemblies and splits.

The metrics are interpenetration rate and total intersection volume, the floating rate, 6-DoF pose accuracy within 10 mm, mate PairF1 read off the final contact graph, and runtime per assembly. The floating rate is the fraction of assemblies whose contact graph is disconnected at a fixed clearance tolerance $\tau$; it is reported for several values of $\tau$, so the conclusion cannot depend on the choice of $\sigma$.

The first decisive check is the mate-quality sweep described above. The prediction is that the barrier and pairwise springs are tied with ground-truth mates and that the gap widens as mate F1 falls. The second check compares the barrier active throughout optimization against the end-only penalty. If the stranding argument is right, the end-only variant will show a much higher floating rate, concentrated on parts whose clearance exceeded about $3\sigma$ early in optimization. The third check sweeps the allowance $\delta$ to locate the point where reduced overlap starts to cause floating, which is the trade-off noted in DM-CFO and Copy-Transform-Paste.

As a consistency check, the method is applied to two-body hand-object refinement on HO-3D and FHB. With only two bodies the barrier reduces to a single attraction weight, so it should roughly reproduce ContactOpt-style behavior, against ContactOpt's reported MPJPE reduction from 57.7 mm to 48.1 mm on HO-3D and CPF's reported 3.71 mm penetration-depth reduction on FHB. Any clear gain or loss there would point to a bug, not a result. The method's real test is assemblies with many parts.