Complementarity-Free Multi-Contact Model

• The model redefines contact mechanics by eliminating complementarity constraints through closed-form, analytic updates.
• It leverages convex optimization and polyhedral friction approximations to achieve differentiable, efficient simulations.
• Empirical results reveal significant speedup and enhanced accuracy in robotic manipulation compared to classical QP/SOCP methods.

A complementarity-free multi-contact model refers to formulations of contact mechanics and robotic multi-body dynamics that encode the physical interaction constraints—such as non-penetration and Coulomb friction—without the use of complementarity conditions or combinatorial hybrid-mode switching. Such models supplant the classical Linear/Mixed Complementarity Problem (LCP/MCP) framework with analytic, smooth, or weakly non-smooth surrogates, typically admitting explicit or closed-form updates for contact forces or velocities. This advance enables efficient, scalable, and differentiable simulation, optimization, and real-time control in contact-rich systems, particularly for dexterous robotic manipulation and locomotion (Jin, 2024).

1. Mathematical Foundations and Model Derivation

The complementarity-free paradigm arises from convex optimization duality applied to the time-stepping equations of multi-contact rigid and compliant bodies. Traditional approaches model contact via KKT conditions for non-penetration and friction, yielding complementarity constraints such as

$$0 \leq \lambda_n \perp \phi(q) \geq 0, \qquad \|\lambda_t\| \leq \mu \lambda_n,$$

where $\lambda_n$ is the normal contact force, $\phi(q)$ is the signed gap, $\lambda_t$ is the tangential force, and $\mu$ is friction coefficient.

The key step introduced in (Jin, 2024) is to replace the dual QP or MCP that would involve per-step solves and combinatorial mode enumeration, by an explicit, closed-form update derived via the dual of the frictional contact QP. After Polyhedral approximation of friction cones (with $n_d$ facets per contact) and stacking $n_c$ contacts, the model reduces, via a diagonal approximation of the dual contact Hessian, to: $$\beta = \max\left(-K(q)\left[\widetilde{J} Q^{-1} b + \widetilde{\phi}\right],\, 0\right)$$

$$v^+ = \frac{1}{h^2} Q^{-1}\left[h b + \widetilde{J}^{\top} \beta\right]$$

where $Q$ is a regularized inertia/impedance matrix, $b$ assembles non-contact wrenches and actuation, $\widetilde{J}$ and $\widetilde{\phi}$ encode the stacked contact constraints, and $K(q)\succ0$ is a diagonal contact-stiffness per constraint. This bypasses the inner-solve, producing velocities and states in analytic or piecewise-smooth form (Jin, 2024).

2. Model Properties: Closed-Form, Differentiability, No Combinatorics

The closed-form structure yields several favorable properties:

• No complementarity or slack variables: Contact activation/deactivation and stick/slip transitions emerge natively as $\max(\cdot,0)$ non-smoothness in (3a), without integer/binary variables, complementarity pairs, or explicit mode switching.
• Differentiability: By mollifying $\max(\cdot,0)$ with a smooth function (SoftPlus with a large scale parameter), the full mapping $(q,u)\mapsto v^+$ becomes smooth and suitable for automatic differentiation. This is essential for gradient-based planning, control, and learning (Jin, 2024).
• Friction Law Satisfaction: By construction, the summed impulses for each contact satisfy the polyhedral approximation to the Coulomb friction cone exactly, ensuring physical validity without post-projection (Jin, 2024).
• Analytic Gradient Availability: All model parameters, including contact stiffnesses, gains, and even geometry, may participate in end-to-end learning or estimation.

3. Hyperparameters and Tuning-Free Operation

Complementarity-free models typically require only a minimal and physically interpretable set of hyperparameters:

• A positive regularization $\epsilon$ for object mass,
• Robot impedance gains $K_r$,
• Number of tangential directions $n_d$ per contact (polyhedral friction cone approximation; $n_d=4$ often suffices),
• Smoothing parameter $\gamma$ for softmax/SoftPlus in smooth model variant,
• A (typically scalar) contact-stiffness $K(q)=kI$.

Experiments demonstrate that a single global value $k\in[10^{-2},10^{1}]$ suffices across a diverse set of manipulation scenarios, with no need for per-task or per-object retuning (Jin, 2024). Differentiable formulation further enables $k(q)$ to be optimized or learned online.

4. Efficiency, Complexity, and Algorithmic Implementation

The elimination of complementarity constraints and cone conditions drastically simplifies per-step computational cost:

Method	Step Complexity	Per-Step Run Time (ms)
QP/SOCP (classical)	$O((n_c n_d)^3)$ (dense, worst case)	30–50
Complementarity-free	$O(n n_c n_d)$ multiplications, $\max$	12–20

In the bar-pushing and dexterous manipulation benchmarks (Jin, 2024), the explicit duality-free update achieves $\sim5\times$ speedup over state-of-the-art QP/SOCP solvers, with convergence of the full MPC loop occurring in $20$ interior-point iterations (vs $50+$ for complementarity-based). This enables real-time contact-implicit MPC rates of $50$–$100$ Hz on commodity CPUs.

A high-level MPC pseudocode with complementarity-free model as the stepwise dynamic update enables efficient, robust nonlinear programming using standard solvers (e.g., CasADi/IPOPT), as no inner or bilevel optimization remains (Jin, 2024).

5. Empirical Performance in Dexterous Manipulation

Quantitative results in (Jin, 2024) establish:

Task Family	Success Rate (%)	Position Error (mm)	Orientation Error (deg)	MPC Rate (Hz)
Fingertip 3D In-Air	>90	7.8	11	50–100
TriFinger In-Hand	97.0	6.7	11.8	~97
Allegro On-Palm	97.6	—	11.5	97
ALL (17 objects)	96.5	7.8	11	50–100

Per-step MPC solve times fall in the $12$–$20$ms range, consistently outperforming complementarity-based variants with higher accuracy and stability (Jin, 2024).

6. Relation to Broader Complementarity-Free Modeling Approaches

A range of alternative formulations—penalty-based contact (Wang et al., 10 Sep 2025), energy-gradient (Fiber Monte Carlo) (Wang et al., 10 Sep 2025), SDF/log-sum-exp based collision and dual-cone models (Yang et al., 2024), unconstrained convex optimization of compliant contacts (via compliance/damping) (Castro et al., 2021), and parallelizable GPU-accelerated dual-cone models (Borse et al., 12 Mar 2026)—inherit the central themes of:

• Closed-form or analytic update for contact impulses/forces,
• No discrete/logic variables or mode selection,
• Structural differentiability,
• Suitability for large-scale simulation, real-time MPC, and learning.

Although some (e.g., (Wang et al., 10 Sep 2025, Haninger et al., 2023)) employ different physical surrogates—such as explicit volumetric energy or compliant Kelvin–Voigt primitives—the principles of decoupling contact computation from solver combinatorics and enabling smooth/efficient pipeline are common.

7. Impact and Significance

Complementarity-free multi-contact models have directly closed the gap between model-based and learning-based robotic manipulation in terms of both raw performance and adaptability. By eliminating computational bottlenecks, introducing tuning-free operation, and aligning modeling assumptions with the needs of real-time optimization and control, they have enabled manipulation, locomotion, and contact-rich planning at previously unattainable scales (Jin, 2024). Empirical results demonstrate state-of-the-art success rates and precision on both simulation and hardware, with explicit throughput and stability advantages over complementarity-based benchmarks.

A plausible implication is expanding adoption across model-predictive control, reinforcement learning with model gradients, large-batch simulation, and differentiable physics pipelines for complex robot-environment interactions. The paradigm represents a fundamental shift in contact modeling for robotics, removing the traditional computational “hybrid curse” associated with combinatorial mode enumeration and unlocks practical, scalable, and generalizable solutions to contact-rich control and optimization.