Contact-Explicit Representation in Modeling

• Contact-explicit representation is a modeling approach that treats contact as a primary computed variable rather than an incidental effect, enabling direct supervision and optimization.
• It encompasses diverse formulations—such as dense affordance fields, binary grids, and continuous fields—applied across domains like robotics, molecular design, and fluid dynamics.
• Empirical results show that explicit contact modeling improves localization, reduces computational overhead, and enhances control in tasks from manipulation to simulation.

Contact-explicit representation denotes a class of modeling strategies in which contact is encoded as a direct object of computation rather than being left as a side effect of pose fitting, generic latent embedding, or uniform message passing. In recent literature, this object has taken the form of a dense affordance field over a camera-frame point cloud, a part-wise binary grid over the MANO hand mesh, an explicit contact field queried on an implicit surface, a residual tactile primitive, a binary contact command for control, or a continuous surface parameter inside trajectory optimization (Xu et al., 7 Jan 2026, Jung et al., 7 May 2026, Lee et al., 13 Jun 2025, Leve et al., 2024). The shared objective is to make contact localizable, verifiable, and reusable by downstream inference, planning, or design modules.

1. Conceptual scope

A concise way to organize the literature is to treat contact-explicit representation as a family of explicit state spaces for interaction.

Domain	Explicit contact variable	Representative formulation
Dense hand contact	Part-wise grids on 103 semantic MANO parts, final $\mathbf{c}\in\{0,1\}^{V}$	ContactPrompt (Jung et al., 7 May 2026)
Extrinsic manipulation	Camera-frame pointwise affordance $a_k\in[-1,1]$	UNIC (Xu et al., 7 Jan 2026)
Hand–object interaction	Object-surface contact, part, and direction maps; attractive/repulsive labels	ContactGen (Liu et al., 2023); ContactGrasp (Brahmbhatt et al., 2019)
Visuo-tactile implicit estimation	Surface-registered contact fields coupled to SDFs	ViTaSCOPE (Lee et al., 13 Jun 2025); NDCF (Merwe et al., 2023)
Contact-conditioned control	Binary contact commands or continuous surface parameters $\phi$	ContactMimic (Li et al., 9 Jul 2026); WBCRM (Leve et al., 2024)
Molecular and physical systems	Residue contact probabilities, real contact area ratio, or explicit geometry constraints	ConTact (Ahmed et al., 20 May 2026); TENG model (Sanglap et al., 1 Apr 2026)

These formulations differ in modality and scale, but they agree on one structural decision: contact is made first-class. In "ConTact," for example, CDR design is explicitly decomposed into learning surface complementarity fingerprints, predicting CDR-antigen contacts, and injecting contact-gated antigen features into the sequence head, rather than conflating interface reasoning and amino-acid selection inside uniform message passing (Ahmed et al., 20 May 2026). In "Neural Deforming Contact Field," the contact patch is represented as a probability field over 3D space and then restricted to the predicted surface, rather than being treated as an auxiliary point set detached from geometry (Merwe et al., 2023).

This suggests that contact-explicitness is less a single data structure than a modeling principle: expose the interaction state in a form that can be supervised, thresholded, optimized, or constrained directly.

2. Surface-registered maps and fields

A large branch of the literature represents contact directly on surfaces or observed scene points. "ContactGrasp" defines an object-centric contact map $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$ on uniformly sampled object-surface points, with $c_i\in\{+1,-1\}$ denoting attractive or repulsive regions; this allows functional intent to be attached to the object rather than to a particular hand pose (Brahmbhatt et al., 2019). "ContactGen" extends the same object-centric idea to a richer tuple $F=(C,P,D)$ over object points, where $C$ is a contact map, $P$ is a one-hot part map over $B=16$ hand parts, and $D$ is a direction map that disambiguates which hand region touches the object and from what direction (Liu et al., 2023). "S$a_k\in[-1,1]$0Contact" makes contact explicit as paired per-point fields $a_k\in[-1,1]$1 over a MANO hand mesh and an object point cloud, with values in $a_k\in[-1,1]$2 discretized into 10 bins for training; the graph-based estimator achieves 2.4× fewer learnable parameters and 2× lower GPU memory cost than PointNet/DGCNN-style baselines (Tse et al., 2022).

A second line of work uses continuous fields. "ViTaSCOPE" defines a unified representation

$a_k\in[-1,1]$3

where $a_k\in[-1,1]$4 is signed distance, $a_k\in[-1,1]$5 is contact probability, and $a_k\in[-1,1]$6 is the tactile shear vector; the contact patch is explicitly

$a_k\in[-1,1]$7

In simulation it reports average in-hand pose errors of about 4.103 mm translation and 1.863^\circ rotation, and in the real world 4.049 mm and 0.907^\circ, while real-world contact-patch CD is around $a_k\in[-1,1]$8 (Lee et al., 13 Jun 2025). "NDCF" adopts the same surface-plus-contact logic for deformable objects, with $a_k\in[-1,1]$9 and contact restricted to the zero level set of the learned SDF. On 298 unseen simulated interactions, NDCF improves deformed geometry CD from 5.746 to 0.910, volumetric IoU from 0.817 to 0.987, and contact patch CD from 36.537 to 22.840 relative to a point-cloud baseline (Merwe et al., 2023).

A third variant attaches contact to observed scene geometry rather than to a canonical object or hand surface. "UNIC" constructs a unified extrinsic contact affordance map on the camera-frame point cloud by converting the nearest distance to annotated contact points into a Gaussian score and then mapping it to $\phi$0. The representation directly supports point, line, patch, and multi-object contact chains such as gripper–object–object–environment, and yields 9.6 mm Chamfer distance on unseen contact locations and 33.4 mm on unseen objects (Xu et al., 7 Jan 2026). For monocular hand-held object reconstruction, "Learning Explicit Contact for Implicit Reconstruction of Hand-held Objects from Monocular Images" predicts part-level contacts $\phi$1 and vertex-level contacts $\phi$2 on the MANO hand, then diffuses them into sparse contact volumes used to condition a DeepSDF-style decoder; with predicted hand meshes on HO3D, it reports $\phi$3 up by 11.8%, $\phi$4 up by 8.4%, and CD reduced by 29.3% relative to IHOI (Hu et al., 2023).

3. Part-based and semantically indexed contact

Part decomposition is a recurring strategy for turning dense contact into a manageable symbolic space. "ContactPrompt" begins from the MANO hand mesh $\phi$5 with $\phi$6, partitions it into $\phi$7 semantic parts aligned with functional anatomy and surface orientation, and organizes each part into a part-wise vertex grid $\phi$8 whose rows are ordered from fingertip to wrist and whose within-row order is left to right in the rendered view. The dense prediction pipeline proceeds through a free-form global reasoning stage, a part stage that outputs a JSON list of contacting parts, and a dense stage that predicts binary grids only on the selected parts; vertices outside the active set are fixed to non-contact. This training-free and zero-shot design outperforms previous supervised methods on the MOW benchmark despite requiring no training (Jung et al., 7 May 2026).

"ContactGen" and "ContactMimic" illustrate two different uses of semantic contact indexing. In ContactGen, the part map $\phi$9 answers which hand part touches each object point, while the direction map resolves ambiguity among grasps compatible with the same contact locations (Liu et al., 2023). In ContactMimic, contact is represented as $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$0 for robot body part $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$1 and object semantic part $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$2, then reduced for policy input to $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$3. The actor therefore receives both reference keypoints and explicit body-level contact commands, enabling contact controllability with the same motion geometry (Li et al., 9 Jul 2026).

An analogous idea appears in protein design. "ConTact" predicts a soft contact map $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$4 over CDR positions, using local antigen aggregation over the $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$5 nearest antigen neighbors and a focal binary cross-entropy loss with ground-truth contacts defined by an 8 Å cutoff. The predicted contact is then reused twice: it gates antigen feature injection into the sequence head and reweights the sequence loss by $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$6. On CHIMERA-Bench, the model reports AAR 0.38, RMSD 1.63 Å, fnat 0.67, DockQ 0.73, and epitope F1 0.79, including about 7% RMSD improvement over the next-best baseline and about 10% epitope F1 over GNN baselines (Ahmed et al., 20 May 2026).

The common effect of these semantically indexed schemes is to replace unconstrained dense inference with staged localization. A plausible implication is that contact-explicit representation often gains reliability by interposing an interpretable contact vocabulary between raw geometry and final actuation or sequence prediction.

4. Multimodal and latent contact representations

Not all contact-explicit representations are maps on a visible surface. "ResTacVLA" reformulates tactile sensing as a residual signal

$\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$7

where $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$8 is the latent tactile state and $\mathbf{c}=\{(p_i,c_i)\}_{i=1}^N$9 is the Cross-Modal Predictor’s visual prediction. The residual is then quantized by a VQ bottleneck into Latent Contact Primitives and gated by visual uncertainty through

$c_i\in\{+1,-1\}$0

This design addresses modality collapse by attenuating visually predictable tactile content and injecting tactile information only when it is surprising. Across five real-world tasks, ResTacVLA reports 62.8% average success versus 28.2% for vision-only $c_i\in\{+1,-1\}$1 and 18.8% for Diffusion Policy without tactile input; replacing VQ with continuous residuals causes a -26.7% degradation, and replacing adaptive gating with fixed always-on tactile fusion causes a -13.3% drop (Zhang et al., 3 Jul 2026).

"Beyond Binary" introduces a physics-grounded tactile representation based on Center-of-Pressure, consisting of a 3D force vector $c_i\in\{+1,-1\}$2 and a 3D contact position $c_i\in\{+1,-1\}$3 in the sensor frame. The taxel-to-CoP mapping uses a Gaussian radial weight and a regularized least-squares solve for $c_i\in\{+1,-1\}$4. In blind peg-in-hole insertion, the full CoP representation achieves 0.78 overall success, compared with 0.67 for force vector only, 0.55 for force magnitude, 0.53 for binary contact, 0.50 for contact position only, 0.48 for raw taxels, and 0.43 for proprioception only; under out-of-distribution peg initializations it reaches 0.63, the best among robot policies (Pan et al., 27 May 2026).

World-model studies show that explicit contact structure can be decisive even when the representation is embedded in latent dynamics. "ContactWorld" compares wrist-view images, front-view images, point clouds, TacRGB, TacDepth, and TacFF within a JEPA-style model. Average planning success rises from 20.7% with wrist-view observations and 22.0% with front-view observations to 32.1% with point clouds, and further to 36.1% when point clouds are combined with tactile force-field representations. For long-horizon planning, PointCloud-only versus PointCloud+TacFF improves from 52.1% → 54.4% at 12 steps, 36.6% → 41.6% at 24 steps, 23.7% → 27.8% at 36 steps, and 16.0% → 20.5% at 48 steps (Zhang et al., 11 Jun 2026).

A related but proprioceptive example appears in legged odometry. "Learning Contact Representation for Leg Odometry" learns a latent code $c_i\in\{+1,-1\}$5 from a denoising autoencoder over a compact 5D kinematic observation and fits a 2-component GMM in latent space so that stance probability directly modulates the ZUPT covariance

$c_i\in\{+1,-1\}$6

On real hardware, the self-supervised DAE GRU achieves 0.900 F1 and DAE CNN 0.868 F1, while DAE CNN gives the best odometry with ATE 17.20 m, AHE 31.90^\circ, RPE trans 49.93%, and FPE 28.66 m (Girgin et al., 3 Jun 2026).

5. Contact as a control, planning, and constraint variable

When contact is embedded in decision-making, explicitness often means that contact itself becomes an input command or an optimization state. In "ContactMimic," the policy

$c_i\in\{+1,-1\}$7

tracks keypoints and binary contact labels simultaneously. Contact-following is shaped by balanced-accuracy or TPR–FPR rewards and by a contact distance reward that pulls intended-contact pairs together and pushes unintended-contact pairs apart. The key training device is trajectory augmentation—contact-label flipping, object removal, and inflated geometry—which breaks the shortcut correlation between keypoints and contact labels, allowing the same trajectory to be executed with contact on or off during deployment (Li et al., 9 Jul 2026).

"Explicit Contact Optimization in Whole-Body Contact-Rich Manipulation" makes contact location a continuous decision variable through a smooth outline parameterization $c_i\in\{+1,-1\}$8 for both robot and object surfaces. The state is augmented with contact parameters $c_i\in\{+1,-1\}$9 and $F=(C,P,D)$0, and in-contact configurations satisfy

$F=(C,P,D)$1

This explicit surface parameterization, combined with hierarchical continuous optimization, yields about 99% fewer iterations and about 96% less planning time on average; reported average planning time drops from about 48 minutes to 1 minute 30 seconds (Leve et al., 2024).

A more abstract control-oriented formulation appears in generalized assembly. "Towards Generalized Robot Assembly through Compliance-Enabled Contact Formations" represents contact not by exact points or frictional states but by contact formations, understood as constraint classes that determine the object’s available DOFs. The manipulation process is a path

$F=(C,P,D)$2

and transitions are detected from variational changes in force/torque signals. Using a compliant robot, the method completes insertion tasks with tolerances <0.25mm without prior knowledge of the exact hole location or its orientation (Morgan et al., 2023).

These formulations show a distinct meaning of contact-explicitness: contact is not merely perceived but commanded, optimized, or maintained as a state variable that shapes feasible motion.

6. Broader scientific uses and recurring empirical patterns

Outside robot perception and control, contact-explicit representation has also become a modeling device in physics, materials, and molecular design. In triboelectric nanogenerator modeling, the explicit object is the real contact area extracted from roughness-resolved finite-element contact analysis and carried into electrostatics through the ratio $F=(C,P,D)$3; the framework matches interference reflection microscopy with absolute deviation in contact area fraction of about 0.5–2.6%, and reports about 17.5% lower $F=(C,P,D)$4 than some analytical predictions at larger separations because fringing reduces the idealized field (Sanglap et al., 1 Apr 2026). In precision reducers, explicit contact geometry is resolved through Circle–Circle, Circle–Curve, and Curve–Curve primitives, with mesh stiffness assembled from bending, shear, foundation, and Hertz contact terms; the predicted torsional stiffness is about $F=(C,P,D)$5, within the literature range $F=(C,P,D)$6, and increasing bearing clearance from $F=(C,P,D)$7 to $F=(C,P,D)$8 raises lost motion by 88% and backlash by 330% while reducing stiffness by only 2.9% (Miao et al., 2 Apr 2026).

In numerical fluid mechanics, explicit contact enters as a discrete non-penetration constraint. "Contact-aware simulations of particulate Stokesian suspensions" augments a boundary-integral Stokes solver with a space-time interference-volume gap function and complementarity condition

$F=(C,P,D)$9

so that the contact force is computed as a Lagrange multiplier rather than by ad hoc repulsion. The method permits at least an order-of-magnitude increase in stable time-step size for dense suspensions, and examples use 16 points per vesicle where about 64 were needed without contact resolution (Lu et al., 2016). In nuclear structure, the HMAMD formulation introduces a Gaussian-weighted high-momentum basis that becomes approximately

$C$0

when $C$1, or exactly an extremely narrow Gaussian with width $C$2. For $C$3H with the AV4' interaction, the contact-explicit basis reduces the number of required basis states from 15 to 8 while reproducing the converged energy, with AMD alone at 15.6 MeV, AMD+Gau at -1.76 MeV, AMD+Dis+Gau at -7.38 MeV, and earlier AMD+Dis at -7.40 MeV (Zhao et al., 2018).

Across these domains, the empirical pattern is consistent. Explicit contact variables tend to appear where latent or purely geometric representations become brittle under occlusion, combinatorial contact modes, stiff interactions, or sparse supervision. This suggests that contact-explicit representation is most useful when interaction structure itself is the bottleneck: the representation is designed so that contact can be named, localized, thresholded, and fed back into the model or solver, instead of being inferred only indirectly after the fact.