Shape-Differentiable Contact Model

• Shape-differentiable contact models are formulations in which outputs like signed distances, witness points, and contact normals vary differentiably with respect to geometric parameters.
• They integrate multiple layers—from analytical SDFs and implicit convex surfaces to manifold generation and contact dynamics—using both explicit and learned proxies.
• These models support applications in calibration, state estimation, and robot co-design while addressing challenges such as smoothing trade-offs and gradient correctness.

Searching arXiv for the cited work and closely related papers to ground the article in current literature. arXiv Search Query: all:"shape-differentiable contact" OR all:"differentiable contact manifold" OR all:"differentiable collision detection" OR id:(Beker et al., 19 Apr 2026) OR id:(Beker et al., 23 Feb 2026) OR id:(Mathew et al., 3 Feb 2026) OR id:(Ye et al., 2024) OR id:(Li et al., 2024) OR id:(Negi et al., 20 Jun 2025) OR id:(Strecke et al., 2021) OR id:(Huang et al., 2022) OR id:(Yang et al., 2023) OR id:(Haninger et al., 2023) OR id:(Chaudet-Dumas, 2022) OR id:(Chaudet-Dumas et al., 2020) OR id:(Lawry et al., 2017) OR id:(Lee et al., 2023) OR id:(Si et al., 2024) OR id:(Zhong et al., 2021) OR id:(Zhong et al., 2022) A shape-differentiable contact model is a contact formulation in which contact-relevant outputs depend differentiably on geometric parameters. Depending on the formulation, those outputs may be signed distances, witness points, contact normals, contact manifolds, contact impulses, predicted contact wrenches, or rollout losses. Recent work spans rigid-body contact kinematics based on analytical signed distance functions (SDFs), differentiable rigid-contact dynamics with low-dimensional geometry parameters, and deformable contact solvers with analytic or directional shape derivatives (Beker et al., 19 Apr 2026, Huang et al., 2022). The literature also shows that the term does not denote a single standardized model: some methods differentiate explicit shape parameters, some expose only geometry-aware surrogates, and some learn a differentiable proxy for a fixed contact manifold rather than differentiating shape itself (Negi et al., 20 Jun 2025).

1. Conceptual scope and recurring formulations

In practice, shape-differentiable contact modeling spans several distinct layers. Some methods differentiate the geometric front end of contact handling—surface representation, collision queries, witness points, and manifold construction. Others differentiate the dynamics layer—contact impulses, wrenches, or time-stepping—with respect to geometry parameters. A third group treats contact as a regularized PDE-constrained mechanics problem and derives shape sensitivities of the state and objective. A fourth group is only adjacent: it uses differentiable contact manifolds or contact features as priors for estimation, without exposing explicit derivatives with respect to shape.

Editor's term: the literature can be read as a set of “layers of shape differentiability.”

Layer	Typical differentiated quantity	Representative papers
Geometry and manifold generation	SDF values, witness points, normals, contact manifolds	(Beker et al., 19 Apr 2026, Beker et al., 23 Feb 2026, Mathew et al., 3 Feb 2026)
Rigid-contact dynamics	Wrenches, impulses, time of impact, rollout loss	(Ye et al., 2024, Li et al., 2024, Strecke et al., 2021)
Deformable contact mechanics	Shape, material, friction, initial conditions	(Huang et al., 2022, Chaudet-Dumas, 2022, Chaudet-Dumas et al., 2020)
Task-specific manifold priors	Projection to a sampled contact manifold	(Negi et al., 20 Jun 2025, Lee et al., 2023)

This suggests that “shape differentiability” should be interpreted with care. A method may be differentiable with respect to pose but not shape, differentiable with respect to a low-dimensional geometric parameter but not an arbitrary surface, or differentiable only after replacing hard contact geometry by a smoothed surrogate. Several papers make this distinction explicit: learned manifold projection for calibration is not a full shape-differentiable mechanics model (Negi et al., 20 Jun 2025), and differentiable rigid simulators with geometry-aware collision handling do not automatically provide end-to-end derivatives with respect to shape parameters (Yang et al., 2023).

2. Geometry representations that expose shape derivatives

The most direct route to shape differentiation is to encode shape by analytical or implicit parameters rather than by feature-selection logic. A prominent example is the extruded plane-superquadric intersection (XPSQ) family, introduced as an analytical primitive class for smoothly differentiable and massively vectorizable collision geometry (Beker et al., 19 Apr 2026). Its foundation is the superquadric, augmented by half-space intersections. A half-space SDF is

$$\phi_N(\mathbf{x}) = \mathbf{x}\cdot \mathbf{n} + h,$$

and smooth constructive solid geometry is implemented with log-sum-exp, for example

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$

An XPSQ sweeps a plane-superquadric intersection along a quadratic spline

$$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$

The resulting SDF is differentiable with respect to rigid pose, spline control points, the orientation field $\mathbf{R}(t)$, scale functions $a_x(t),a_y(t),a_z(t)$, shape exponents $\epsilon_1(t),\epsilon_2(t)$, and half-space coefficients $\mathbf{P}(t)$. The same paper reports that a cup is represented with only three XPSQ primitives, whereas convex decomposition yields 36 primitives, and that an armadillo is represented by 18 SQs (Beker et al., 19 Apr 2026).

A second line of work regularizes geometry itself into strictly convex implicit surfaces. iDCOL represents colliding bodies by strictly convex implicit surfaces $\phi:\mathbb{R}^3\to\mathbb{R}$, motivated by the observation that flats, edges, and undefined curvature create intrinsic non-uniqueness in contact location and normal (Mathew et al., 3 Feb 2026). The method replaces exact sharp geometry by smooth strictly convex approximations, including LogSumExp smooth-max blends,

$$\phi_{\beta}(\mathbf y)=\frac{1}{\beta}\log\sum_{j=1}^m \exp(\beta c_j(\mathbf y)),$$

and superellipsoidal families with $\varepsilon$-regularization to avoid vanishing curvature. This does not by itself provide a full shape-optimization pipeline, but it turns the geometry-to-contact map into a smooth implicit problem whose solution can be differentiated analytically (Mathew et al., 3 Feb 2026).

DiffSDFSim adopts a different implicit route: each object is represented by an SDF

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$0

or, in a learned shape space,

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$1

Surface normals are extracted from SDF gradients,

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$2

and the paper explicitly proposes differentiation with respect to shape parameters, including primitive dimensions and latent codes $$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$3 (Strecke et al., 2021). This representation supports nonconvex geometry and allows shape to affect both collision geometry and inertia.

A more combinatorial but still differentiable representation appears in SDRS, which models each link as a union of convex polyhedra

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$4

Here the design variables are the convex-hull vertices themselves. The representation is intended to remain differentiable under substantial geometry and topology changes, because it avoids explicit contact-feature instantiation and instead builds contact around separating planes between convex hulls (Ye et al., 2024).

Another geometry-aware formulation, aimed primarily at pose estimation, represents convex shape through a prescribed support function

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$5

with support points given by

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$6

Since the support function depends directly on vertices $$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$7, the construction is structurally compatible with shape differentiation even though the demonstrated optimization variables are poses rather than shape parameters (Lee et al., 2023).

3. Contact kinematics and manifold generation

Shape-differentiable contact requires more than a differentiable surface representation; it also requires differentiable generation of contact points, normals, and manifold structure. The XPSQ work addresses precisely this front end. It assumes one object is represented as an SDF and the other as a mesh, then constructs a geometric contact manifold in two stages: edge-SDF intersection and per-face aggregation (Beker et al., 19 Apr 2026). For a mesh edge

$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$8

the paper avoids differentiating through an argmin and instead uses a fixed-iteration sphere-tracing-style procedure from both endpoints; three iterations were empirically sufficient. Each triangular face yields six potential contact points, from which the method computes depths, activity indicators, normals, and contact Jacobians, then fuses them with softmax weights based on depth. A key design decision is that normals and Jacobians are fused directly rather than by first fusing positions. The paper is explicit that this is not yet a full contact model in the dynamics sense; it is a differentiable manifold generator intended to feed a later contact dynamics layer (Beker et al., 19 Apr 2026).

A related but more directly manifold-oriented framework introduces smoothly differentiable and efficiently vectorizable contact manifold generation through vertex-SDF and differentiable edge-edge routines (Beker et al., 23 Feb 2026). Geometry is represented simultaneously by a mesh and a smooth analytical SDF. Contact candidates are selected by soft top-$$\phi_\cup(\mathbf{x}) = -\mathrm{LSE}([-\phi_1(\mathbf{x}),-\phi_2(\mathbf{x})]), \qquad \phi_\cap(\mathbf{x}) = \mathrm{LSE}([\phi_1(\mathbf{x}),\phi_2(\mathbf{x})]).$$9 operations, and edge-edge witness points are computed by a regularized, branch-softened box-constrained quadratic program rather than by discrete active-set logic. The method then assigns signed distances and signed normals and constructs a smooth activity indicator $$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$0 from soft feasibility, penetration, nearest-neighbor, and clashing tests. The paper reports a significant speedup against MuJoCo XLA in its benchmark (Beker et al., 23 Feb 2026).

iDCOL formulates contact kinematics itself as a fixed-size nonlinear system. Using scaled convex bodies, it solves

$$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$1

and derives witness points, a signed gap-like quantity,

$$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$2

and the contact normal from the optimal solution (Mathew et al., 3 Feb 2026). With the KKT residual $$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$3, the derivative is obtained by the Implicit Function Theorem: $$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$4 This is a contact-kinematics model rather than a full contact solver, but it supplies analytically differentiated distances, witness points, normals, and contact wrenches for downstream dynamics (Mathew et al., 3 Feb 2026).

DiffSDFSim also addresses kinematics, but with a different emphasis. Contact points are extracted from differentiable meshes produced by marching cubes over the SDF, then refined by a Frank–Wolfe search over mesh triangles. Penetration is read directly from the SDF,

$$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$5

and the paired point is approximated by stepping along the gradient,

$$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$6

Its most distinctive contribution is the time-of-contact differential. For impact contacts it defines a collision-time constraint $$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$7 and differentiates

$$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$8

so that shape changes influence not only contact Jacobians but also when the impact occurs (Strecke et al., 2021). This closes a gradient path that fixed-step rigid contact models usually miss.

4. Contact dynamics and force laws

Once contact geometry is available, different papers choose different force and time-stepping layers. In rigid-body simulation, Jade models contact as a linear complementarity problem (LCP) and uses continuous collision detection with time-of-impact backtracking to remain intersection-free (Yang et al., 2023). Its one-step dynamics are built around

$$\mathbf{p}(t) = (1-t)^2\mathbf{p}_1 + 2t(1-t)\mathbf{p}_2 + t^2 \mathbf{p}_3, \qquad t\in[0,1].$$9

and the paper derives gradients with respect to state, controls, inertial parameters, and timestep. However, it does not develop explicit derivatives with respect to shape parameters themselves; geometry enters through CCD, contact Jacobians, and event handling rather than through a formal shape calculus (Yang et al., 2023).

A direct extension of this line estimates low-dimensional geometric uncertainty by differentiating predicted contact wrenches with respect to geometry parameters $\mathbf{R}(t)$0 (Li et al., 2024). The simulator output is

$\mathbf{R}(t)$1

and the geometry-sensitive terms include contact points, normals, Jacobians, and the time of impact. The paper derives

$\mathbf{R}(t)$2

but requires the user to provide $\mathbf{R}(t)$3 and $\mathbf{R}(t)$4 for each URDF. This yields a differentiable rigid-contact pipeline with respect to low-dimensional geometric parameters, not a general automatic shape-differentiable collision engine (Li et al., 2024).

SDRS removes explicit feature selection from the contact law by introducing a separating plane

$\mathbf{R}(t)$5

for each pair of contacting convex polyhedra and defining the contact potential through an auxiliary optimization,

$\mathbf{R}(t)$6

The plane is interpreted as a zero-mass auxiliary entity whose state is determined by stationarity. The paper proves that the resulting contact potential is globally twice-differentiable in $\mathbf{R}(t)$7, diverges as distance tends to zero, vanishes when the hulls are sufficiently separated, and preserves linear and angular momentum (Ye et al., 2024). This is an explicitly shape-differentiable rigid-contact model, but it is a penalty/barrier model over convex hulls rather than a classical feature-based contact manifold.

In deformable contact, the most complete shape-differentiable dynamic framework uses Incremental Potential Contact (IPC). The normal contact energy is

$\mathbf{R}(t)$8

with deformed positions $\mathbf{R}(t)$9, barrier stiffness $a_x(t),a_y(t),a_z(t)$0, and area weights $a_x(t),a_y(t),a_z(t)$1 (Huang et al., 2022). Friction is regularized by

$a_x(t),a_y(t),a_z(t)$2

where $a_x(t),a_y(t),a_z(t)$3 smooths the stick–slip transition. The paper then derives analytic adjoints for shape, material, friction, and initial conditions. This is a genuine shape-differentiable contact mechanics framework, but it achieves differentiability by replacing exact unilateral contact and Coulomb friction with barrier and smoothing constructions (Huang et al., 2022).

Static contact mechanics offers a stricter mathematical view. For small-strain elasticity with contact against a rigid foundation, shape differentiation is derived for a penalty formulation and for an augmented Lagrangian formulation, but only through directional derivatives of the nonsmooth projection operators

$a_x(t),a_y(t),a_z(t)$4

The corresponding papers prove shape differentiability only under additional assumptions, notably that weak contact and weak sticking sets have measure zero (Chaudet-Dumas, 2022, Chaudet-Dumas et al., 2020). A related finite-strain level-set/XFEM formulation with large sliding and frictionless contact supplies adjoint sensitivities for shape and topology optimization, but the contact law still relies on active-set updates and other piecewise-smooth mechanisms (Lawry et al., 2017).

A simpler parametric approach is to replace hard contact by always-active compliant primitives. One example models each contact as

$a_x(t),a_y(t),a_z(t)$5

with geometry parameters $a_x(t),a_y(t),a_z(t)$6 and $a_x(t),a_y(t),a_z(t)$7 as differentiable variables (Haninger et al., 2023). This is smoothly differentiable with respect to contact location and rest position, but it is not a full object-shape model.

5. Applications: estimation, calibration, control, and co-design

Shape-differentiable contact models are used in several application regimes. In-space kinematic calibration is one example. A learned contact manifold model is trained offline as a metric projection onto the set of peg-hole contact poses and used online to estimate thermal strain $a_x(t),a_y(t),a_z(t)$8 and encoder bias $a_x(t),a_y(t),a_z(t)$9 from encoder readings collected only at contact times (Negi et al., 20 Jun 2025). The contact signal is binary,

$\epsilon_1(t),\epsilon_2(t)$0

and the learned manifold operates in a 6D Euclidean pose parameterization. Across 20 trials, using 3000 contact observations collected at 300 Hz, the paper reports strain MAE $\epsilon_1(t),\epsilon_2(t)$1 and encoder bias MAE $\epsilon_1(t),\epsilon_2(t)$2, corresponding to a 2.6-fold reduction in link strain error and a 3.37-fold reduction in joint encoder bias error, with sub-mm and sub-degree end-effector pose accuracy (Negi et al., 20 Jun 2025). This is an effective differentiable contact-manifold model for calibration, but not a full explicit shape-differentiable mechanics model.

Stable placement under geometric uncertainty provides another use case. A differentiable contact dynamics model built on Jade is used to estimate uncertain object pose, object shape, or environment geometry from force-torque discrepancies during placement (Li et al., 2024). The geometry parameter $\epsilon_1(t),\epsilon_2(t)$3 may represent in-hand pose offsets, object wall heights, environment pillar deviations, or saucer location, and the system maintains a belief over multiple hypotheses to mitigate local minima. The reported average computation time is 3.36 s per action, so the method is not real-time, but it demonstrates that gradients through contact wrenches can support online geometric uncertainty estimation (Li et al., 2024).

DiffSDFSim is directly targeted at shape inference. In a bouncing-sphere experiment without gravity, the proposed time-of-contact differential yields mean final radius error $\epsilon_1(t),\epsilon_2(t)$4 and max error $\epsilon_1(t),\epsilon_2(t)$5; with gravity, mean final radius error improves from $\epsilon_1(t),\epsilon_2(t)$6 without the time-of-contact term to $\epsilon_1(t),\epsilon_2(t)$7 with it (Strecke et al., 2021). In learned shape spaces, average Chamfer distance improves from $\epsilon_1(t),\epsilon_2(t)$8 to $\epsilon_1(t),\epsilon_2(t)$9, and a separate shape-from-inertia experiment reduces average Chamfer distance from $\mathbf{P}(t)$0 to $\mathbf{P}(t)$1. The paper also reports a real tennis-ball reconstruction in which the estimated radius improves from $\mathbf{P}(t)$2 cm after initial fitting to $\mathbf{P}(t)$3 cm, compared with a ground-truth $\mathbf{P}(t)$4 cm (Strecke et al., 2021).

Shape-differentiable contact also appears in robot co-design. SDRS is explicitly designed so that both robot shapes and control movements can be optimized simultaneously, using convex-polyhedron link geometry and separating-plane contact (Ye et al., 2024). A related but lower-dimensional tradition appears in differentiable compliant contact primitives for estimation and MPC, where online estimation of contact rest positions $\mathbf{P}(t)$5 and local contact locations $\mathbf{P}(t)$6 improves force tracking in planar and pivoting tasks (Haninger et al., 2023).

On the geometric-feature side, uncertain pose estimation during contact uses differentiable support-function contact features inside a bi-level estimator (Lee et al., 2023). The lower-level problem estimates contact forces under friction-cone constraints, while the upper-level problem optimizes pose. In peg-in-hole experiments, the paper reports runtimes of $\mathbf{P}(t)$7–$\mathbf{P}(t)$8 ms for a rectangular peg and $\mathbf{P}(t)$9–$\phi:\mathbb{R}^3\to\mathbb{R}$0 ms for a hexagonal peg, compared with $\phi:\mathbb{R}^3\to\mathbb{R}$1–$\phi:\mathbb{R}^3\to\mathbb{R}$2 ms and $\phi:\mathbb{R}^3\to\mathbb{R}$3–$\phi:\mathbb{R}^3\to\mathbb{R}$4 ms for a 25-particle filter, respectively (Lee et al., 2023). These are pose-estimation results rather than shape-optimization results, but they demonstrate the practical value of differentiable contact-feature geometry.

A tactile analogue appears in DIFFTACTILE, a differentiable tactile simulator with FEM-based elastomer mechanics and penalty-based contact

$\phi:\mathbb{R}^3\to\mathbb{R}$5

Its contact forces depend directly on signed distance, normal direction, and deformed surface geometry, and the entire simulator is implemented in Taichi with automatic differentiation (Si et al., 2024). The paper does not provide explicit shape derivatives, but it is an example of a contact-rich differentiable mechanics stack in which geometry-sensitive gradients are available in principle.

6. Limitations, controversies, and open problems

The field is unified less by a single mathematical framework than by a recurring set of compromises. One compromise concerns exactness versus smoothness. Analytical SDF unions, barrier energies, smooth friction ramps, softened top-$\phi:\mathbb{R}^3\to\mathbb{R}$6, and regularized witness-point solves make gradients available, but the resulting derivatives are often derivatives of a smoothed surrogate rather than of the original hard contact law. Several papers state this openly: XPSQ-based manifold generation uses soft branch merging, soft min/max, and soft activity indicators (Beker et al., 19 Apr 2026); IPC uses a positive-separation barrier and smoothed friction rather than exact zero-gap Signorini contact and Coulomb friction (Huang et al., 2022); support-function contact features rely on a smoothed support function rather than the exact nonsmooth convex hull support map (Lee et al., 2023).

A second controversy concerns gradient correctness. A comparative study of differentiable contact simulators reports that gradients with respect to position, velocity, and control are often not correct, even in simple rigid-body collision problems, and that discarding time of impact in velocity-impulse models yields completely wrong gradients with respect to position; making the timestep smaller does not fix this issue (Zhong et al., 2022). This has a direct implication for shape differentiation. Since shape parameters typically influence the same event structure—gap, contact normal, active feature, and impact time—any failure in state gradients is a warning about geometry gradients as well. DiffSDFSim’s explicit time-of-contact differential can be read as one response to this problem (Strecke et al., 2021).

A third limitation is partial scope. Some papers address only the geometric front end. XPSQ-based manifold generation is explicit that it solves geometry representation and manifold construction, not friction laws, complementarity conditions, impulsive resolution, or time integration (Beker et al., 19 Apr 2026). Geometry-aware rigid simulators such as Jade remain geometry-dependent without becoming fully shape-differentiable; they handle continuous collision detection and event-time differentiation but do not derive $\phi:\mathbb{R}^3\to\mathbb{R}$7, $\phi:\mathbb{R}^3\to\mathbb{R}$8, or $\phi:\mathbb{R}^3\to\mathbb{R}$9 (Yang et al., 2023). Learned manifold projection for in-space calibration is geometry-dependent but object-specific and not explicitly differentiable with respect to peg or hole shape parameters (Negi et al., 20 Jun 2025).

A fourth limitation is manual modeling burden. Some methods require the user to specify derivatives of contact features with respect to geometry parameters for each model (Li et al., 2024), or to choose and fit a dual representation consisting of mesh plus SDF primitives (Beker et al., 23 Feb 2026). Others lack an automatic route from raw meshes to the smooth primitive family that contact differentiation assumes; the XPSQ paper explicitly states that automatic decomposition of arbitrary meshes into XPSQ primitives remains future work (Beker et al., 19 Apr 2026).

A fifth limitation concerns nonsmooth residual events that survive smoothing. Static contact shape optimization still requires measure-zero assumptions on weak contact and weak sticking sets to obtain full shape differentiability (Chaudet-Dumas, 2022, Chaudet-Dumas et al., 2020). IPC-based deformable simulation remains piecewise smooth in practice because active contact pairs, collision search results, and remeshing can change discretely (Huang et al., 2022). These caveats indicate that global $$\phi_{\beta}(\mathbf y)=\frac{1}{\beta}\log\sum_{j=1}^m \exp(\beta c_j(\mathbf y)),$$0 or $$\phi_{\beta}(\mathbf y)=\frac{1}{\beta}\log\sum_{j=1}^m \exp(\beta c_j(\mathbf y)),$$1 contact-through-shape theorems are rare.

Taken together, these results suggest that a shape-differentiable contact model is best viewed not as a single method but as an architectural goal: a simulator or estimator in which geometry is parameterized in a differentiable way, contact kinematics are generated without branch-heavy feature logic, and the remaining force, friction, and time-stepping layers preserve useful gradients under geometry variation. The strongest current formulations either regularize geometry itself, as in strictly convex implicit contact kinematics (Mathew et al., 3 Feb 2026), regularize contact mechanics, as in IPC-based deformable simulation (Huang et al., 2022), or re-engineer manifold generation to avoid combinatorial discontinuities (Beker et al., 23 Feb 2026, Beker et al., 19 Apr 2026). The main open problem is to combine these properties—shape expressivity, contact fidelity, computational efficiency, and trustworthy gradients—without relying on task-specific simplifications or heavy smoothing.