Certified Empty Space for Open-Surface Neural Distance Fields

Neural distance fields are small networks that take a 3D point and return its distance to a surface. They are increasingly used to plan robot motion and detect collisions around real objects. Many real objects are open sheets: leaves, garments, car body panels, wires. For these, the classic signed representation, which needs a well-defined inside and outside, breaks down. The field has moved to unsigned distance fields (UDFs), which predict only a non-negative distance. Nobody can currently prove anything about what these networks say. Every meshing or collision pipeline over a UDF samples a grid and hopes nothing thin slipped between the samples.

Formal neural network verification tools compute provable bounds on a network's outputs over a whole input region, and they have recently been extended beyond piecewise-linear activations to tanh, sine and products. We propose to use them to certify, region by region, that a UDF network contains no surface there. The certificate pairs two ingredients. The first is a provable lower bound on the distance value. The second is a provable upper bound on the size of the field's gradient. The gradient bound turns one evaluated distance into a certified empty ball, which is the provable version of the sphere-tracing step graphics already uses. We use this to answer a question neither community has posed: which activation functions and field designs give open-surface geometry that is both accurate and cheap to certify?

Core Claim

The defining difficulty of UDFs is that the true gradient flips direction by 180° when crossing the surface (Neural Unsigned Distance Fields for Implicit Function Learning; MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks). This breaks any certificate that reasons about gradient direction. It does not break a certificate that reasons about gradient magnitude. A well-trained UDF approximately satisfies the eikonal property, meaning its gradient has norm close to 1, so a certified norm bound should be close to 1 and therefore tight.

Concretely, let $f$ be the trained field, $\tau \ge 0$ the threshold below which we treat a point as surface, $x_0$ a query point with certified $f(x_0) \ge d$, and $L$ a certified upper bound on $\|\nabla f\|$ over the ball $B(x_0, R)$ of radius $R$. The mean value theorem along straight segments inside the convex ball then gives

$$ f(y) > \tau \quad \text{for all } y \text{ with } \|y - x_0\| < \min\big(R,\ (d-\tau)/L\big). $$

So one bounded value and one bounded gradient norm certify a ball nearly as large as the distance itself. A value-only certificate has to bound $f$ over the whole box. That forces the verifier to branch heavily near the surface, where many neurons are ambiguous and the curvature of smooth activations hurts most.

The claim has two parts. First, combining value and gradient-norm certificates gives much larger certified empty volume, and a thinner certified band enclosing all of the surface, than value-only bound propagation at equal compute. Second, the activation function and the way the field is parameterized change certifiability by large factors that reconstruction metrics do not reveal. A model can top the accuracy table and be nearly uncertifiable.

Relation to Prior Work

The verification side supplies the method. Formal Verification for Neural Networks with General Nonlinearities via Branch-and-Bound (GenBaB) adds branch-and-bound to bound propagation for sine, tanh, GeLU and multiplications. Branch-and-bound means splitting an input region into smaller pieces when the bounds are too loose to decide. GenBaB places split points using precomputed lookup tables. It raised verified instances on a 4×100 sine network on CIFAR-10 from 4% to 60%. It certifies only classification margins and power-grid constraints, never fields over 3D coordinates, and we reuse its value bounds unchanged.

Scalable Verification of Neural Control Barrier Functions Using Linear Bound Propagation propagates linear bounds through network Jacobians and refines the state space with simplices. It verified the Darboux benchmark in about 145 seconds where the dReal solver timed out. It assumes smooth barrier functions because it needs full Jacobian bounds. We need only a norm bound on the Jacobian, which stays well defined across the UDF kink. This is the one technical change that lets the tool apply to unsigned fields.

The exact region-enumeration approach of Efficient Verification and Falsification of ReLU Neural Barrier Certificates is a constraint rather than a method. It needed about 31,000 seconds on a 6-12-12-1 ReLU network, which rules out exact enumeration for distance fields with hundreds of hidden units.

The geometry side supplies the models and the motivation. NDF (Neural Unsigned Distance Fields for Implicit Function Learning) introduced ReLU UDFs that can represent interior structures, such as a car's steering wheel inside its shell. Two later methods resolve the sign ambiguity on a grid without guarantees: MeshUDF: Fast and Differentiable Meshing of Unsigned Distance Field Networks votes on pseudo-signs from gradient flips, and Neural Surface Detection for Unsigned Distance Fields classifies each grid cell with a small MLP. Their known failure modes, inconsistent votes and cracks between cells, are exactly what a continuous certificate would expose.

Metric—Phase Fields: Decoupling Distance and Sign for Thin-Structure Reconstruction from Unoriented Point Clouds (MPF) writes the field as $r(x)\tanh(\beta\theta(x)) + \theta(x)$, which avoids vanishing gradients at the surface. Its tanh and product terms are precisely the operators GenBaB handles, so MPF is a natural stress test for certifiability.

CATNIPS: Collision Avoidance Through Neural Implicit Probabilistic Scenes is the closest safety-oriented work. It gives probabilistic voxel bounds on radiance-field density rather than deterministic guarantees on a distance network. We do not claim to replace it, because the two methods certify different representations.

Research Program

We first build the certifier. Over a domain cube we run an adaptive octree. Each cell is attacked first by a ball certificate from its center: GenBaB value bounds give $d$, and Jacobian bound propagation gives a norm bound $L$ over the enclosing ball. The cell falls back to a value-only box bound if the ball does not cover it, and is split if both fail. Cells that end up certified as empty are removed. The cells that remain, down to a minimum size, form a certified band that provably contains every point the network places within $\tau$ of zero.

Branching for the norm bound uses GenBaB's precomputed split points for the activation. One addition targets the kink region. When the norm bound fails, we split the region along the network's approximate gradient direction rather than along coordinate axes. The intent is that each child piece lies mostly on one side of the surface, where gradient directions are consistent.

The certifier is sound with respect to the network, not the physical object. A robot that plans on the network's zero set gets a provable statement about that model. Separately, we measure empirically how far the model's zero set sits from the true geometry.

Second, we run an activation and parameterization study at matched parameter count and training budget. The activations are ReLU (as in NDF), softplus, tanh, GeLU and sine. The parameterizations are a plain UDF with a clamped output, MPF, and a ReLU UDF trained with an eikonal-plus-Jacobian-norm penalty. For each model we record two numbers: reconstruction accuracy, and certification tightness per unit of verifier compute. Together these trace an accuracy–certifiability frontier.

Third, we test whether a cheap training change moves models along that frontier. The change is penalizing the bound-propagated gradient norm over sampled balls, analogous to certified training for classifiers. We check whether it buys certifiability without losing fidelity on thin structures.

Baselines

The naive reference point is what every current pipeline does. We evaluate the network on a dense grid at $128^3$ and $256^3$, as in MeshUDF and Neural Surface Detection, and declare cells empty when sampled values exceed $\tau$. It provides no guarantee. We include it to measure how often it is wrong: we take every cell it calls empty and search it with gradient-based falsification for points below $\tau$. Any found point is a demonstrated miss.

The strongest existing method is value-only certification with GenBaB inside the same octree, using the same per-cell time limit, the same total GPU budget, and the same minimum cell size. It is the most capable general-nonlinearity verifier in the file and requires no new ideas to apply. No reported result was found on Lacuna for GenBaB, or any verifier, applied to 3D distance fields. Its closest reported figure is the CIFAR-10 sine result of 60% verified instances, which is not comparable to a geometric task. We beat it if, at equal wall-clock time, the combined certifier certifies a clearly larger fraction of the truly empty volume and produces a thinner band on ShapeNet cars and garments. We set the bar at a band at least a factor of two thinner at the median.

A Lipschitz-only baseline evaluates $f$ at points and uses a global gradient bound from the product of layer spectral norms. This is the classical graphics approach of certifying with a Lipschitz constant.

The key ablation removes the Jacobian certificate from our pipeline while keeping the octree and the gradient-aligned splitting. A second ablation replaces gradient-aligned splitting with axis-aligned splitting. Together they isolate which ingredient produces any gain.

Why not just…?

Why not just sample densely and use a global Lipschitz constant? Spectral-norm products overestimate the local gradient of deep networks, often by large factors. Certified balls then shrink toward zero and the method degenerates into dense sampling. The Lipschitz-only baseline measures this directly. If its certified volume matches ours on shallow NDF-sized decoders, we will report that local bounds are only needed for deeper or sine-based models.

Why not just make the network 1-Lipschitz by construction? This is the most serious threat. A network that is 1-Lipschitz by design makes the ball certificate free, needing no Jacobian bound propagation at all. We are not aware of any such open-surface UDF in this file. The open question is whether such architectures keep thin-sheet fidelity: MPF reports that unsuitable designs bulge or fragment on thin structures, with a Chamfer distance of 5.591 for SDF baselines against 0.672 for MPF. We will therefore train a 1-Lipschitz UDF variant as an extra arm of the study. If it matches unconstrained models on Chamfer distance, the honest conclusion is that architecture beats verification, and the project's contribution narrows to certifying existing unconstrained models.

Why not just watertight the shapes and verify a signed field with mature ReLU tools? The literature already shows the cost on the geometry side. NDF documents that watertight methods lose interior structures, and MeshUDF documents double-walled artifacts from inflation. We still include an inflated-SDF arm on garments. This measures whether the certifiability gain of signed fields is worth their geometric error, quantified as the Chamfer distance against the original open surface.

Proposed Experiments

Training and evaluation follow the benchmarks in the file. We use ShapeNet cars with interior structures under NDF's protocol, where NDF reports a Chamfer-$L_2$ of 0.05. We use Multi-Garment Net clothing, for which no reported number was found on Lacuna. We use ABC CAD at $128^3$, where Neural Surface Detection reports a median Chamfer distance of 2.52 against 2.97 for MeshUDF. And we use the MPF thin-structure models, where MPF reports a Chamfer distance of 0.672.

For every model we report five quantities: reconstruction Chamfer distance; certified empty fraction of the volume known to be empty from ground truth; median and 95th-percentile thickness of the certified band; verifier time; and falsification-found misses, which must be zero for certified methods and are counted for grid sampling.

A soundness check comes first. Ten million random and adversarially optimized points inside certified-empty regions must all satisfy $f > \tau$. Any violation is treated as a bug, not a result.

Three outcomes would settle the claims. First, the combined certifier should beat value-only GenBaB on band thickness at matched compute, and the Jacobian ablation should account for most of that gap. Second, the activation study should show rank reversals, meaning a model with better Chamfer distance but worse certifiability, especially between sine and softplus or tanh networks and for MPF's product term. If certifiability simply tracks accuracy, the second claim fails. Third, on garments and car panels, grid sampling at $128^3$ should have nonzero demonstrated misses near thin structures. That would show the guarantee catches failures that ordinary practice cannot see.

As a downstream demonstration, we plan certified collision-free corridors through ShapeNet car interiors and around garments. We compare corridor width and planning time against grid-thresholded planning on the same networks.