Optimization-Based Contact Models

• Optimization-based contact models are mathematical frameworks that formulate contact dynamics as optimization problems enforcing non-penetration and friction laws.
• Recent complementarity-free and closed-form models improve differentiability, parallelization, and integration into trajectory optimization and learning frameworks.
• Embedding these models in planning and control enhances scalability, real-time performance, and accuracy in simulating complex, contact-rich physical systems.

Optimization-based contact models are a foundational paradigm in computational rigid-body and soft-body dynamics for simulating, planning, and controlling contact-rich physical systems. These models formalize non-penetration and friction laws as optimization or variational problems—typically convex quadratic, conic, or energy-based programs—enabling the prediction of contact forces, impulses, and subsequent state evolution in a principled and often computationally efficient manner. Historically, complementarity constraints were viewed as necessary for enforcing contact exclusivity (non-penetration and frictional sticking/sliding), but recent advances have yielded a variety of differentiable, closed-form, and fully parallelizable models that eliminate the need for complementarity, thereby accelerating computation and enabling seamless embedding in robotic optimal control, learning, and real-time simulation frameworks.

1. Classical Formulation and Limitations

Traditional optimization-based contact models for rigid and compliant multibody dynamics are constructed by discretizing the equations of motion and introducing variables for contact impulses $\lambda$ alongside complementarity constraints enforcing non-penetration and Coulomb friction. The standard approach, typified by the Anitescu–Potra or Stewart–Trinkle methods, leads to a mixed linear complementarity problem (LCP) or a second-order cone program (SOCP) at every simulation or planning time step: $$\text{minimize}_v \quad \frac{1}{2} h^2 v^\top Q v - h b^\top v \quad \text{subject to} \quad J_{i,j}(q) v + \phi_i/h \geq 0, \quad \forall (i,j)$$ where $Q$ encodes inertia and stiffness, $b$ collects forces, $J_{i,j}$ are friction-cone faces, and $\phi_i$ are contact distances. The complementarity conditions introduce non-smoothness and combinatorial complexity, hindering global differentiation, regularization, and parallelizability. These limitations have been circumvented in recent work by relaxing, regularizing, or eliminating complementarity altogether, leading to complementarity-free, optimization-based contact models with superior computational and differentiation properties (Yang et al., 2024, Jin, 2024, Borse et al., 12 Mar 2026, Castro et al., 2021, Lee et al., 2022).

2. Complementarity-Free and Closed-Form Analytical Models

Complementarity-free optimization-based contact models reconstruct the fundamentals of contact dynamics via regularized, algebraic, or energy-based approaches, yielding closed-form or easily parallelizable evaluations. Methods span a spectrum:

• Spring-damper/Impedance formulations: Each contact is modeled via linear or nonlinear force–penetration laws without discrete mode switching. Friction is enforced via saturation, regularization, or soft cone projections, ensuring differentiability (Chatzinikolaidis et al., 2020, Han et al., 2024, Haninger et al., 2023).
• Primal/Dual Cone Approximations: Coulomb friction cones and non-penetration half-spaces are replaced by polyhedral or soft approximations; in the “dual-cone” perspective, velocity inequalities replace force constraints and closed-form clamped corrections are applied along cone-facet directions (Borse et al., 12 Mar 2026, Jin, 2024).
• Signed Distance Function (SDF) Layers: Contact geometry and dynamics are encoded as differentiable SDFs parameterized via log-sum-exp “soft-max” smoothings, producing analytic velocity and state updates through explicit gradient flows (Yang et al., 2024).
• Energy-based Contact via Overlap Volume: Contact is modeled by potential energies depending on the measure of overlap between domains (e.g., $E = k (V_c)^m$), with forces computed as gradients. This generalizes to arbitrary geometries and material models and bypasses mesh conformity and master–slave choices (Wang et al., 10 Sep 2025).
• Surrogate Nodalization and Diagonalization: Dynamics are reformulated so that contact resolution decouples at the node or axis level, enabling parallel, analytically tractable local solves with proven consistency to the original coupled dynamics (Lee et al., 2022).

These approaches permit state updates, sensitivity analysis, and gradient-based learning without nested QP or LCP solves and with explicit, smooth mappings from system state to contact forces.

3. Formulations: SDF, Impedance Models, and Convex Energy Methods

Specific classes of optimization-based contact models are distinguished by their mathematical structure and computational properties:

• Signed Distance Function (SDF) contact models utilize convex support functions and log-sum-exp aggregations to encode collision detection (C-SDF) and multi-contact time stepping (D-SDF), providing smooth, differentiable approximations to both geometric and frictional constraints. The ContactSDF framework defines

$$\mathrm{csdf}_G(x_{\text{query}}) = \frac{1}{\sigma} \mathrm{LSE} \left( 0, \mathrm{LSE}_i\{\sigma (n_i^\top x_{\text{query}} + b_i)\} \right)$$

and analogous closed-form D-SDF layers, collectively yielding an explicit, end-to-end-differentiable multi-contact model (Yang et al., 2024).

• Impedance-style prediction–correction models employ time-discretized, dual-cone impedance corrections, decoupled across contact points and cone facets, yielding per-facet clamped impulse updates with

$$\lambda^{(j)}_s = \left( -K^{(j)}_s (s^{(j)}_s\, dt + \tilde{\phi}^{(j)}) - D^{(j)}_s s^{(j)}_s \right)_+$$

These admit massive GPU parallelization and scale nearly linearly with the number of contacts (Borse et al., 12 Mar 2026).

• Unconstrained convex energy minimization eliminates all contact constraints via Moreau-envelope regularization and projection operators, leading to a globally convergent, differentiable, single-stage minimize-problem for all contact points:

$$\min_{y \in \mathbb{R}^{3m}} f(y) = \frac{1}{2} (b + B \operatorname{Proj}_C(y))^\top A^{-1}(b + B \operatorname{Proj}_C(y)) + \sum_i \psi_i(y_i)$$

(Castro et al., 2021). The $\operatorname{Proj}_C$ operator analytically projects onto the friction cone and normal non-negativity.

• Fiber Monte Carlo (FMC) energy-based contact models replace discrete contact constraints with the gradient of an energy functional defined by overlap volume, computed via fiber-based integral estimators, ensuring differentiability for complex geometries and mesh independence (Wang et al., 10 Sep 2025).

4. Embedding in Planning, Control, and Model Learning

Optimization-based contact models are naturally suited for embedding in direct transcription trajectory optimization (TO), model predictive control (MPC), and system identification via gradient-based learning. With complementarity-free or analytic contact formulations, the resultant optimal control problem reduces to a smooth or piecewise-smooth nonlinear program: $$\min_{q_0,\ldots,q_N} \; \ell_f(q_N) + h \sum_{k=1}^{N-1} [ \ell(q_k) + c(\tau(q_{k-1}, q_k, q_{k+1})) ]$$ subject to explicit, differentiable state transitions given by the contact model (Chatzinikolaidis et al., 2020, Yang et al., 2024, Jin, 2024, Haninger et al., 2023). SDF-based or dual-cone models permit automatic differentiation through the entire stack, supporting system identification and model adaptation from real or simulated data.

Adaptive or online learning of model parameters, such as mass, stiffness, damping, or SDF sharpness, can be naturally cast as minimizing state prediction error over experimental rollouts, leveraging gradients provided by the closed-form contact update rules (Yang et al., 2024). In practice, these advances shrink the reality gap and enable rapid experimental adaptation.

5. Scalability, Performance, and Benchmarks

Complementarity-free optimization-based contact models scale favorably with system complexity. Analytical or decoupled update rules permit:

• Linear or near-linear scaling in contact count via decoupled facet or contact-point solves (Borse et al., 12 Mar 2026, Lee et al., 2022).
• High-frequency control: Models such as ContactSDF and dual-cone/impedance solvers achieve real-time MPC at 30–100 Hz and simulation in the tens of kilohertz for underactuated systems with modest computational hardware (Yang et al., 2024, Borse et al., 12 Mar 2026, Jin, 2024).
• Massive parallelism: GPU-based implementations show up to $2\times$ higher throughput and substantially reduced per-step latency in dense contact scenes, e.g., in-hand robotic manipulation benchmarks (Borse et al., 12 Mar 2026).

Empirical evaluation indicates that complementarity-free models not only match the physical fidelity of mainstream QP/SOCP-based engines but also support end-to-end differentiability and more robust solver convergence. For instance, in dexterous manipulation tasks, these models achieve >96% success rates and sub-centimeter accuracy, whereas classical implicit-contact MPCs are limited to lower update rates and greater computational cost (Jin, 2024).

6. Assumptions, Model Choices, and Limitations

Key modeling assumptions and trade-offs inherent to optimization-based contact models include:

• Contact compliance: Most approaches introduce explicit stiffness/damping parameters or energy regularization, controlling penetration tolerance and numerical conditioning (Chatzinikolaidis et al., 2020, Castro et al., 2021, Haninger et al., 2023).
• Polyhedral friction cones: Friction cones are often discretized into $n_d$ directions for computational tractability. Increasing $n_d$ improves friction accuracy at increased computational cost (Jin, 2024, Yang et al., 2024).
• Convexity and differentiability: Analytical or energy-based relaxations trade off strict enforcement of non-penetration and stiction for the sake of global convexity and differentiability, which may introduce small penetrations or approximation errors on hard or high-frequency contacts (Castro et al., 2021, Han et al., 2024).
• Quasi-static and explicit dynamics regimes: Some SDF and spring-based approaches assume negligible inertial terms, making them best suited for slow, manipulation-dominated regimes (Yang et al., 2024).
• Parameter tuning: Contact stiffness, friction, and smoothing parameters require tuning either from empirical data or by system identification; improper selection may affect physical plausibility or solver performance (Jin, 2024).

Nonetheless, data-driven fitting, analytical approximations, and hardware-in-the-loop adaptation have been shown to mitigate these limitations effectively in real systems.

7. Contemporary Directions and Research Outlook

Contemporary research in optimization-based contact modeling focuses on unifying efficiency, accuracy, and differentiability for contact-rich robotic systems in challenging settings:

• Automated system identification and adaptive learning of contact model parameters using trajectory rollouts and real-world feedback (Yang et al., 2024, Haninger et al., 2023).
• Scalable, real-time simulators and physics engines, including GPU-accelerated, parallel frameworks compatible with high-fidelity planning and reinforcement learning pipelines (Borse et al., 12 Mar 2026, Lee et al., 2022).
• Dense, multi-modal, and arbitrary-geometry contact via energy- or overlap-based constructs, eliminating legacy requirements of mesh conformity or master–slave rules (Wang et al., 10 Sep 2025).
• Robustness to unmodeled dynamics, frictional uncertainty, and complex nonconvex geometries through compositional primitives, SDF smoothing, and fiber-based integral estimators (Wang et al., 10 Sep 2025, Yang et al., 2024).

A plausible implication is that future model architectures will continue to integrate analytic, differentiable contact modeling with parallel hardware acceleration and on-the-fly adaptation, thus supporting scalable learning, planning, and control in robotics, computer graphics, and computational mechanics.

---

Key References:

• ContactSDF: Signed Distance Functions as Multi-Contact Models for Dexterous Manipulation (Yang et al., 2024)
• Complementarity-Free Multi-Contact Modeling and Optimization for Dexterous Manipulation (Jin, 2024)
• ComFree-Sim: A GPU-Parallelized Analytical Contact Physics Engine for Scalable Contact-Rich Robotics Simulation and Control (Borse et al., 12 Mar 2026)
• Contact-Implicit Trajectory Optimization using an Analytically Solvable Contact Model for Locomotion on Variable Ground (Chatzinikolaidis et al., 2020)
• Solving contact problems using Fiber Monte Carlo (Wang et al., 10 Sep 2025)
• Online Multi-Contact Feedback Model Predictive Control for Interactive Robotic Tasks (Han et al., 2024)
• Large-Dimensional Multibody Dynamics Simulation Using Contact Nodalization and Diagonalization (Lee et al., 2022)
• Differentiable Compliant Contact Primitives for Estimation and Model Predictive Control (Haninger et al., 2023)
• An Unconstrained Convex Formulation of Compliant Contact (Castro et al., 2021)