Hunt–Crossley Model for Nonlinear Contact Dynamics

• The Hunt–Crossley model is a dynamic law that fuses a nonlinear elastic (power-law) response with depth-dependent viscous dissipation to simulate contact interactions.
• Advanced estimation techniques, including log transformation with recursive least squares, EKF, and ARMCMC, enable real-time parameter identification despite hybrid system challenges.
• Widely adopted in soft robotics, haptics, and biomedical applications, the model serves as a benchmark for studying hybrid and nonlinear dynamical systems.

The Hunt–Crossley model provides a canonical dynamic law for characterizing contact interactions between an indenter and a compliant surface. Its structure fuses a nonlinear elastic (power-law) response with depth-dependent viscous dissipation, supporting accurate modeling in contexts such as soft robotics, haptics, and biomedical indentation. The model is widely used both as a forward physical law and as a challenging benchmark for real-time parameter estimation in hybrid and nonlinear dynamical systems.

1. Mathematical Formulation and Symbol Definitions

The original Hunt–Crossley contact-force law is:

$$f_e\bigl(x(t)\bigr)\;=\; \begin{cases} K_e\,x(t)^{\,p}\;+\;B_e\,x(t)^{\,p}\,\dot x(t) & x(t)\ge 0 \quad\text{(contact mode)}, \ 0 & x(t)<0 \quad\text{(free-motion mode)}. \end{cases} \tag{1}$$

Where:

Symbol	Interpretation
$x(t)$	Penetration depth (indentation of indenter into material)
$\dot x(t)$	Rate of change of penetration (velocity)
$K_e$	Non-linear stiffness coefficient (elastic term)
$B_e$	Non-linear damping coefficient (viscous term)
$p$	Nonlinearity exponent ($1 \leq p \leq 2$, "Hertzian" if $p = 3/2$)
$f_e$	Resulting normal force

This law exhibits two discrete regimes: "contact" (when $x\ge0$), governed by the nonlinear force terms, and "free motion" (when $x<0$), where the force is strictly zero. This hybrid (piecewise) structure introduces a sharp switching manifold at $x=0$.

2. Model Reduction and Dimensionality Reduction

Online identification of all three core Hunt–Crossley parameters ($K_e, B_e, p$) can be technically challenging due to nonlinearity and identifiability issues. If the contact geometry is a known axisymmetric indenter (e.g., a sphere of radius $R$), dimensionality reduction methods can be applied to analytically map the 3D contact problem onto an equivalent 1D system. Under the parabolic (quasi-Hertzian) approximation, the force law becomes:

$$F_M(d, \dot d) = \kappa\,d^{3/2} + \lambda\,d^{1/2}\,\dot d \qquad (d \geq 0) \tag{2}$$

with

$$\kappa = \frac{16G}{3}\sqrt{R}, \qquad \lambda = 8\eta\sqrt{R}$$

where $d$ is the indentation depth, $G$ is the shear modulus, and $\eta$ the viscosity of the soft medium. This reduction eliminates the need for explicit estimation of the exponent, fixing it at $3/2$ for spherical geometry (Beber et al., 2024).

3. Assumptions and Approximations

The Hunt–Crossley formulation makes several assumptions for both physical modeling and system identification:

• Hybrid regime: Contact force is piecewise: zero in free motion, nonlinear in contact.
• No friction, adhesion, or hysteresis: Only normal force is modeled; tangential effects and irreversible processes are omitted.
• Measurement noise: Additive, independent, and typically modeled as Gaussian (mean $\mu_\nu$, variance $\sigma_\nu$).
• Small-damping approximation: For linear-in-parameters recasting, the log approximation is used:

$$B_e/K_e\,\dot x \ll 1 \implies \log(1+B_e/K_e\,\dot x) \approx B_e/K_e\,\dot x$$

(used for schemes such as recursive least squares) (Agand et al., 2022).

• Constant material properties: Parameters $K_e, B_e$ (or $\kappa,\lambda$) are spatially uniform during a contact episode.

These approximations enable model simplification, tractable parameter estimation, and embedding within filtering frameworks.

4. Hybrid and Identification Challenges

The Hunt–Crossley model exhibits hybrid, multi-modal behavior due to its piecewise definition:

• Discontinuous switching: At $x = 0$ (or $d = 0$), the force abruptly switches from zero to nonlinear contact, invalidating smooth parametric evolution.
• Parameter reset: All "effective" parameters become zero in free motion; on re-contact, they resume previous or new values.
• Parameter estimation impact: Recursive least squares, classical Kalman filters, and particle filters can struggle due to their reliance on smooth parametric drift and unimodal distributions. Particle filters not leveraging domain physics may suffer particle depletion at abrupt resets.
• The Hunt–Crossley law is thus used as a canonical benchmark for identification methods in systems with abrupt mode transitions and nonlinear, non-LIP dynamics (Agand et al., 2022, Beber et al., 2024).

5. Parameter Estimation and Identification Algorithms

Several strategies have been implemented for real-time identification of Hunt–Crossley parameters:

5.1 Linear-in-Parameters Transformation

Under small-damping ($B_e/K_e\,\dot x\ll1$), a log transform yields:

$$\log f_e \approx p\,\log x + \log K_e + \frac{B_e}{K_e}\dot x$$

or

$\log f_e \approx U(t)^\top \theta$

where:

$U(t)$	$\theta$
$$\begin{bmatrix}1\ \dot x(t)\ \log x(t)\end{bmatrix}$$	$$\begin{bmatrix}\log K_e\ B_e/K_e\ p\end{bmatrix}$$

Enabling recursive least squares and other LIP-based methods (Agand et al., 2022).

5.2 Recursive and Probabilistic Methods

• Adaptive Recursive MCMC (ARMCMC): Uses the true nonlinear (original) Hunt–Crossley law, leveraging a temporal forgetting factor for efficient, online, hybrid-aware sampling. ARMCMC alternates between "reinforcement" (updating prior from latest posterior) and "modification" (resetting prior), based on a model-mismatch index $\zeta^t$. The Maximum-A-Posteriori (MAP) point estimate is extracted per update (Agand et al., 2022).
• Extended Kalman Filter (EKF): The reduced/quasi-Hertzian version is embedded in a discrete-time EKF for estimating contact force and tissue parameters online, supporting both sensor-based and sensorless implementations (via Cartesian robot impedance control). The state vector includes indentation, its rate, elastic and viscous parameters; the system is updated with kinematic or robot-control measurements (Beber et al., 2024).

6. Practical Applications and Experimental Insights

• Robotics and Haptics: Enables real-time estimation of contact force, depth, stiffness ($\kappa$), and viscosity ($\lambda$), supporting applications such as palpation, soft-tissue characterization, and dynamic manipulation.
• Biomedical estimation: Sensorless EKF implementations reconstruct indentation force and tissue parameters using robot-internal measurements, obviating the need for dedicated force sensors—useful for non-invasive diagnostic tools. Experiments with silicone phantoms (including "tumor" inclusions) show the estimator converges to ground-truth parameters within 10–20% in ~5s; stiffness jumps reveal the presence of inclusions (Beber et al., 2024).
• Benchmarking identification algorithms: The model’s hybrid and nonlinear nature makes it a stringent testbed for Bayesian methods, as demonstrated with ARMCMC, which achieves much lower parameter tracking error and reduced sample requirements compared to conventional MCMC and recursive least squares (Agand et al., 2022).

7. Implementation Details and Algorithmic Settings

• Data acquisition: Sampling interval $T = 1\,\mathrm{ms}$; data processed in blocks of $N_s=100$ ($T_s=100\,\mathrm{ms}$ per update).
• Parameter priors: Non-informative Gaussian prior $\mathcal{N}(1,\,0.1^2)$ on transformed parameters $\theta = [\log K_e, B_e/K_e, p]$.
• MCMC control: Accuracy $\epsilon=0.01$, reliability $\delta=0.90$; typical ARMCMC update requires ~6,000 samples per batch (vs 15,000 for vanilla MCMC).
• EKF state/measurement: State $[d, \dot d, \kappa, \lambda]^\top$; measurement can be either direct force sensor output or robot velocity. Estimates both mechanical parameters and physical indentation states contemporaneously.
• Contact assumptions: Indenter shape and material constants known; tissue isotropic, incompressible; no out-of-plane motion during contact.

Both (Agand et al., 2022) and (Beber et al., 2024) highlight that adaptation and identification are performed at the filter/proposal level, with the physical Hunt–Crossley force law itself left unchanged.

---

References:

• "Online Probabilistic Model Identification using Adaptive Recursive MCMC" (Agand et al., 2022)
• "Towards Robotised Palpation for Cancer Detection through Online Tissue Viscoelastic Characterisation with a Collaborative Robotic Arm" (Beber et al., 2024)