betaFeedback

WorksComputer Science

CROPS: Cosserat-Rod-Based Co-Design of Soft Slender Programmable Material Systems for Biomedical Applications

Abstract

Slender structures made of stimuli-responsive materials represent an emerging class of programmable material systems, with applications in soft robotics, minimally invasive surgery, wearable electronics, and deployable structures. In biomedical applications, the on-demand design of slender structures, encompassing both their material layout and actuation, is essential to harnessing their large deformation and multiphysics actuation to meet patient-specific design requirements. However, owing to geometric and material nonlinearities, their structure-actuation-response relationship is highly nonintuitive and computationally demanding. This often makes trial-and-error and conventional design approaches challenging for addressing patient-specific needs. To this end, we present a Cosserat-rod-based design optimization framework for the autonomous design of programmable slender structures. Leveraging the Cosserat rod representation, the end-to-end differentiable framework is demonstrated on two case studies: (i) design of swellable stents that morph to the shape of carotid vessels; (ii) co-design of the remnant magnetization and the applied field of magnetically-active catheters to reduce the force exerted on the patient-specific vessel wall during insertion. Both self-contact and external contact with surrounding structures are accounted for. We consider two actuation mechanism, namely the swelling of gel-elastomer composites and the magnetic actuation of hard-magnetic composites. The proposed framework presents an autonomous design platform of soft slender material systems that support large deformations and multiphysics actuation toward personalized biomedical design.

Highlights

  • We present a differentiable design optimization framework for stimuli-responsive slender structures.

  • Cosserat rod modeling of slender structures accommodates large deformations, multiphysics actuation, and contact.

  • Spatially resolved material distributions and actuation variables are jointly optimized for swelling in gel-elastomer composites and magnetic actuation in hard-magnetic soft materials.

  • Patient-specific stent and catheter designs demonstrate the capabilities of the proposed framework.

Overview of the proposed framework, CROPS

Figure 1. Overview of the proposed framework, CROPS (Cosserat Rod Optimization of Programmable Slender structures). Given (A) a slender structure, represented as a Cosserat rod, the framework optimizes (B) the material distribution along the length and across the cross-section of the rod, together with the applied actuation. The framework is demonstrated on two devices designed for (C) patient-specific vessel anatomy. In (D) a magnetically actuated catheter, the remnant magnetization and the applied magnetic field are co-designed, reducing the force exerted on the vessel wall during insertion in comparison to an unmagnetized catheter. In (E) a shape-conforming stent, the distribution of the hydrogel and elastomer phases is optimized such that a stent that is straight when dry morphs to the shape of the vessel upon swelling.

Introduction

Programmable material systems (PMS) form a class of material systems that can adapt their shapes or properties in response to external stimuli, enabling dynamic changes in shape, motion, properties, or functions mediated by the stimuli. PMS have been developed for a range of functions, including mechanical actuation and locomotion [1, 2, 3, 4], tuning of mechanical properties [5, 6], and shape programming and reconfiguration [7, 8, 9]. Across these applications, functional behavior often emerges from nonlinear coupling among spatial material distribution, geometry, and actuation. Optimizing these variables independently or fixing some a priori may overlook design combinations that achieve better mechanical performance, motivating their co-design [10]. Co-design can therefore expand the range of achievable responses while better satisfying geometric and functional requirements; however, exploring this coupled design space remains computationally demanding because it requires repeated evaluations of nonlinear mechanical responses throughout design optimization.

For PMS composed of active slender structures, Cosserat rod theory [11] offers a promising approach to reducing the computational burden of co-design. Slender structures serve as versatile building blocks for PMS, enabling complex three-dimensional shape changes in applications ranging from shape-morphing lattice structures [12, 7] to steerable soft continuum robots [4, 13]. In conducting the analysis and design of such structures, Cosserat rod theory represents slender structures using centerline positions and cross-sectional orientations along their length [11]. Such formulations have been applied to modeling large deformations [14, 11], swelling-driven shape changes [15], and interactions involving contact and friction [16, 17]. Cosserat rod-based models can reduce the computational cost of repeated design evaluations involving nonlinear finite element analyses, complex boundary conditions, and demanding functional requirements across states, making iterative material–geometry–actuation co-design more tractable than the standard three-dimensional finite element methods developed for general purposes.

A computationally efficient co-design framework holds significant promise for biomedical devices such as catheters and stents. Their designs must accommodate patient-specific anatomical constraints while satisfying coupled mechanical, fluidic, and other functional requirements, subject to stringent safety and regulatory standards. As such, existing biomedical design approaches have explored different subsets of this coupled design space, with many focusing on programming mechanical performances. For stents, gradient-based optimization has been used to tailor mechanical stiffness and flexibility through strut dimensions, junction positions, and periodic unit-cell topology [18, 19, 20]. For magnetically actuated catheters, evolutionary optimization has been utilized to tailor axial magnetic-particle distributions under prescribed geometry and actuation conditions [13]. More integrated approaches have explored sequential optimization of catheter geometry and material distribution [21], as well as joint optimization of magnetization profiles and applied magnetic fields [22, 23]. Despite these advances, existing catheter co-design approaches remain

The Python source code is available at github.com/ideal-nu/crops limited to coarse axial segments or discrete material choices, rather than continuous three-dimensional vector profiles and cross-sectional layouts under contact. Gradient-free optimization often suffers from slow convergence and weak scalability as the number of design variables in the joint design space increases.

Addressing these limitations requires both spatially resolved material parameterizations and optimization methods that remain tractable as the number of design variables grows. Topology optimization has enabled the design of spatially heterogeneous material distributions in hard-magnetic soft materials [24], reprogrammable magneto-mechanical structures [6], and multi-material gel–elastomer systems [25]. Differentiable simulation further enables sensitivity-based optimization of coupled material and actuation variables, as demonstrated in magnetic soft robots undergoing large deformation and contact using the material point method [1]. For slender structures, gradient based approaches have been utilized to optimize centerline shapes and parameters within prescribed cross-sectional architectures using rod models with linear elastic sectional laws [26, 27]. However, extending these approaches to locally parameterized material distributions along the rod and across its cross-section requires design sensitivities that account for nonlinear coupling among material response, actuation, and deformation. Cross sectional homogenization of heterogeneous material responses incorporates local material behavior into the reduced rod model without independently resolving the three-dimensional displacement field. By differentiating through this homogenization and the nonlinear Cosserat rod solution, including contact, we demonstrate that our framework can support spatially resolved material–actuation co-design while accounting for nonlinear deformation and contact.

Building upon our previous work on differentiable multi-material topology optimization [25], we develop CROPS, a differentiable Cosserat rod-based optimization framework for programmable slender material systems.

The main contributions are threefold:

  1. Differentiable Multiphysics Rod Formulation: We develop a differentiable Cosserat rod formulation that incorporates fiber-resolved cross-sectional integration to model heterogeneous material responses under swelling and magnetic actuation. The framework further incorporates regularized contact mechanics and implicit differentiation to compute sensitivities of deformation and contact forces with respect to material and actuation variables.

  1. Spatial Material–Actuation Co-Design: We establish a gradient-based optimization framework that accommodates continuous-valued material distributions along the rod axis and across its cross-section, enabling joint optimization of spatial material properties, magnetization profiles, and applied actuation.

  1. Demonstrations for Patient-Specific Biomedical Design: We demonstrate the framework through the inverse design of gel-elastomer stents to achieve prescribed deployed geometries and co-design of magnetically actuated catheters to minimize wall-contact forces given patient-specific insertion trajectories.

Method

Overview

In this study, we consider the design optimization of programmable slender structures, in which the material distribution within the structure is optimized to obtain prescribed responses under actuation. We adopt a geometrically exact Cosserat rod [28, 11] to represent the slender structure. In addition to resolving the deformation through the centerline, we retain a material description over the cross-section. In particular, we consider two forms of actuation (Figure 2). The first form of actuation is a gel-elastomer composite in which an active gel swells upon exposure to a solvent while a passive elastomer resists and redirects the deformation. The other one is a hard-magnetic composite in which particles carrying a remnant magnetization embedded in a soft matrix develop moments under an applied magnetic field. In both cases, the spatial distribution of material encodes the mapping from stimulus to deformation.

Actuation modes considered in this work

Figure 2. Actuation modes considered in this work. (a) Reference configuration of the rod with the prescribed Dirichlet ΓD\Gamma_D and Neumann ΓN\Gamma_N boundaries. (b) Swelling actuation of a gel-elastomer composite. Upon immersion, the hydrogel phase absorbs solvent and expands while the elastomer phase remains passive, and the resulting differential expansion drives the deformation. (c) Magnetic actuation of a hard-magnetic composite. Under the applied field Ba\mathbf{B}^{a}, the embedded remnant magnetization develops moments that deform the rod. The remnant direction is parameterized by its azimuth α\alpha and polar angle β\beta in the global frame.

We begin by assuming an unstressed reference configuration of the centerline with prescribed arc length and specified cross-sectional orientations along the curve. Furthermore, we assume a prescribed set of candidate material phases and their properties, for instance the gel and elastomer phases of the swelling system. Additionally, we assume prescribed fixtures, forces, and moments. We also assume prescribed environmental conditions: (i) for swelling, these include the chemical potential, temperature, and molar volume of the solvent bath; (ii) for magnetic actuation, these include the bounds on the magnitudes of the remnant magnetization and the applied magnetic field. Where relevant, the geometry of confining surroundings and obstacles is also prescribed. Our design goal is to compute the material distribution, and where relevant the applied magnetic field, for which the actuated structure attains a prescribed target shape or mechanical output, while imposing the design constraints.

We now consider the rod’s centerline position x(s)∈R3\mathbf{x}(s) \in\mathbb{R}^{3} parameterized by the reference arc length s∈[0,L]s \in[0,L], together with an orthonormal director triad {d1(s),d2(s),d3(s)}\{\mathbf{d}_{1}(s),\mathbf{d}_{2}(s),\mathbf{d}_{3}(s)\}. Here, d1\mathbf{d}_{1} and d2\mathbf{d}_{2} span the cross-sectional plane while d3\mathbf{d}_{3} is normal to it. The triad encodes the cross-sectional orientation, admitting bending, twist, shear, and stretch. As illustrated in Figure 3, the centerline is discretized into NeN_e elements over NnN_n nodes, with nodal positions xi\mathbf{x}_i, nodal triads {d1i,d2i,d3i}\{\mathbf{d}_1^i,\mathbf{d}_2^i,\mathbf{d}_3^i\}, and element ee spanning nodes ii and i+1i+1 with reference length Δse\Delta s_e. Each node carries three translational and three rotational degrees of freedom (DOF), q(i)=[δxi,δyi,δzi,θxi,θyi,θzi]T\mathbf{q}^{(i)} = [\delta_x^i,\delta_y^i,\delta_z^i,\theta_x^i,\theta_y^i,\theta_z^i]^{\mathsf{T}}, assembled into the global state vector q∈R6Nn\mathbf{q} \in\mathbb{R}^{6N_n}. Here, δi\boldsymbol{\delta}^i are displacements from the reference nodal positions x0i\mathbf{x}_0^i and θi\boldsymbol{\theta}^i are rotation vectors that carry the reference directors onto the current ones, both resolved in the global frame. For convenience, we collect the directors into the rotation matrix D∈SO(3)\mathbf{D} \in SO(3), whose columns are d1\mathbf{d}_1, d2\mathbf{d}_2, and d3\mathbf{d}_3, and which maps material frame components onto the global frame, so that D=exp⁡(θ^)D0\mathbf{D} = \exp(\hat{\boldsymbol{\theta}})\mathbf{D}_0, with θ^\hat{\boldsymbol{\theta}} the skew-symmetric matrix of θ\boldsymbol{\theta} and D0\mathbf{D}_0 the reference director frame.

Figure 3. Discretized representation of the rod. The centerline is divided into elements of reference length Δse\Delta s_e spanning nodes xi\mathbf{x}_i and xi+1\mathbf{x}_{i+1}, each node carrying a director triad {d1i,d2i,d3i}\{\mathbf{d}_1^i,\mathbf{d}_2^i,\mathbf{d}_3^i\}. The DOF are illustrated at an arbitrary node mm, where the displacements δx\delta_x, δy\delta_y, δz\delta_z are measured from the reference position xm0\mathbf{x}_m^0 and the rotations θx\theta_x, θy\theta_y, θz\theta_z carry the reference directors onto the current ones. The reference configuration is shown in gray and all global quantities in purple.

Furthermore, we let the material vary both along the length and within the cross-section. Towards this, we resolve the cross-section into discrete fibers at prescribed offsets from the centerline, as illustrated in Figure 4. A fiber ff is located at a fixed material offset ξ(f)=(ξ1(f),ξ2(f))\boldsymbol{\xi}^{(f)} = (\xi_1^{(f)},\xi_2^{(f)}) in the (d1,d2)(\mathbf{d}_1,\mathbf{d}_2) basis. We refer to the portion of fiber ff within element ee as a fiber segment, where each fiber segment carries its own material. It is the material assigned to the fiber segments that we seek to design.

Fiber-resolved representation of the cross-section

Figure 4. Fiber-resolved representation of the cross-section, illustrated for a gel-elastomer composite. Left, the fiber tracks swept along the length of the rod and colored by material phase, with cross-sections shown at selected stations. Right, a single cross-section of node mm, where a fiber ff is located at the fixed material offset ξ(f)=(ξ1(f),ξ2(f))\boldsymbol{\xi}^{(f)} = (\xi_1^{(f)},\xi_2^{(f)}) in the (d1m,d2m)(\boldsymbol{d}_1^m,\boldsymbol{d}_2^m) basis.

We consider the finite deformation elastic response of the structure together with the swelling or magnetic actuation, the contact with the surroundings and with itself, and any applied mechanical loads. The equilibrium configuration is obtained by rendering the total potential energy stationary with respect to the degrees of freedom. We adopt gradient-based optimization, with the sensitivities computed via automatic differentiation (AD) through our end-to-end differentiable pipeline.

The remainder of this section is organized as follows. Section 2.2 presents the kinematics and the elastic response. Section 2.3 and Section 2.4 describe the swelling and magnetic actuation, respectively, and Section 2.5 the contact and applied loading. The design parametrization and the optimization problem are presented in Section 2.6, followed by the sensitivity analysis in Section 2.7 and the complete algorithm in Section 2.8.

Kinematics and Elastic Response

Having established the state representation, we now define the strain measures that characterize the deformation of the rod. We present a condensed account here, and refer the reader to [28, 11] for details.

We begin by defining the curvature κ(s)\boldsymbol{\kappa}(s), which describes the rate of change of the director triad along the arc length ss through ∂sda=κ×da\partial_s\boldsymbol{d}_a = \boldsymbol{\kappa} \times\boldsymbol{d}_a for a=1,2,3a = 1, 2, 3. Resolved in the material frame, κ1\kappa_1 and κ2\kappa_2 measure bending about the two cross-sectional axes, while κ3\kappa_3 measures twist. Similarly, the centerline tangent resolved in the material frame defines the stretch and shear vector ν(s)\boldsymbol{\nu}(s), with components νa=∂sx⋅da\nu_a = \partial_s\boldsymbol{x} \cdot\boldsymbol{d}_a. Here, ν1\nu_1 and ν2\nu_2 measure the transverse shear of the cross-section, while ν3\nu_3 measures the axial stretch. Together, (ν(s),κ(s))(\boldsymbol{\nu}(s), \boldsymbol{\kappa}(s)) characterize the centerline deformation. Finally, we note that the reference strains (νo,κo)(\boldsymbol{\nu}^o,\boldsymbol{\kappa}^o) follow from the prescribed reference centerline and directors, and are in general nonzero.

Having described the kinematics of the centerline, we now turn to the individual fibers. Recall that a fiber f=1,…,Nff = 1,\ldots,N_f is located at a fixed material offset ξ(f)=(ξ1(f),ξ2(f))\boldsymbol{\xi}^{(f)} = (\xi_1^{(f)},\xi_2^{(f)}) in the (d1,d2)(\boldsymbol{d}_1,\boldsymbol{d}_2) basis (Figure 4). Fiber attributes vary both along the arc length and across the cross-section, and are evaluated at the fiber segment of element ee and fiber ff as (⋅)ef(\cdot)_{ef}. We suppress these indices wherever no ambiguity arises. Furthermore, assuming the cross-section deforms rigidly in its own plane, a fiber occupies the position x(s,ξ(f))=x(s)+ξ1(f)d1(s)+ξ2(f)d2(s)\boldsymbol{x}(s,\boldsymbol{\xi}^{(f)}) = \boldsymbol{x}(s) + \xi_1^{(f)}\boldsymbol{d}_1(s) + \xi_2^{(f)}\boldsymbol{d}_2(s). Observe that the fiber positions are determined entirely by the centerline and the director triad. In other words, the fiber bundle introduces no DOF beyond those in qq. Differentiating along the arc length and resolving in the material frame yields the total fiber stretch λ=ν3−ξ1κ2+ξ2κ1\lambda= \nu_3 - \xi_1\kappa_2 + \xi_2\kappa_1 and the two fiber shears γ1=ν1−ξ2κ3\gamma_1 = \nu_1 - \xi_2\kappa_3 and γ2=ν2+ξ1κ3\gamma_2 = \nu_2 + \xi_1\kappa_3. We denote the reference stretch and shear components by λ0\lambda_0, γ1,0\gamma_{1,0}, and γ2,0\gamma_{2,0}, respectively. Then, the local deformation gradient at the fiber, resolved in the material frame, can be written as:

F=[λt0η10λtη200λ],(1)\mathbf{F} = \begin{bmatrix} \lambda_t & 0 & \eta_1 \\ 0 & \lambda_t & \eta_2 \\ 0 & 0 & \lambda \end{bmatrix}, \tag*{(1)}

where λ=λ~/λ~0\lambda= \tilde{\lambda}/\tilde{\lambda}_0, η1=(γ1−γ1,0)/λ~0\eta_1 = (\gamma_1 - \gamma_{1,0})/\tilde{\lambda}_0, and η2=(γ2−γ2,0)/λ~0\eta_2 = (\gamma_2 - \gamma_{2,0})/\tilde{\lambda}_0. Furthermore, λt\lambda_t is the transverse stretch, taken as isotropic within the cross-sectional plane. Following from the definition of volume ratio J=det⁡F=λt2λJ = \det\mathbf{F} = \lambda_t^2\lambda, we have

λt=Jλ.\lambda_t = \sqrt{\frac{J}{\lambda}}.

Having established the fiber kinematics, we now specify the elastic response. We adopt a neo-Hookean strain energy density

ψela=G2(I1−3−2ln⁡J),(2)\psi_{\mathrm{ela}} = \frac{G}{2}\left(I_1 - 3 - 2\ln J\right), \tag*{(2)}

where GG is the shear modulus and I1=tr⁡(FTF)I_1 = \operatorname{tr}(\mathbf{F}^{\mathsf{T}}\mathbf{F}) is the first invariant of the right Cauchy-Green tensor. Furthermore, we assume molecular incompressibility, such that the volume ratio JJ is prescribed by the swelling state of the fiber, with J=1J = 1 in the absence of swelling (§ 3.3). The elastic response is one of several contributions to the free energy density of a fiber, which in general we write as

ψ=ψela+ψgel+ψmag,(3)\psi= \psi_{\mathrm{ela}} + \psi_{\mathrm{gel}} + \psi_{\mathrm{mag}}, \tag*{(3)}

where the swelling and magnetic contributions, ψgel\psi_{\mathrm{gel}} and ψmag\psi_{\mathrm{mag}}, are introduced in § 3.3 and § 3.4 respectively, and vanish where the corresponding actuation is absent.

The internal energy follows by integrating this density over the cross-section and along the arc length. The cross-sectional integral is evaluated as a weighted sum over the fibers, with the reference fiber area AefA_{ef} serving as the quadrature weight, so that

Πint(q)=∑e=1NeΔse∑f=1NfAef ψef.(4)\Pi_{\mathrm{int}}(\mathbf{q}) = \sum_{e=1}^{N_e} \Delta s_e \sum_{f=1}^{N_f} A_{ef}\,\psi_{ef}. \tag*{(4)}

The total potential energy comprises this internal energy together with the contact energy Πcon\Pi_{\mathrm{con}} and the work of the applied loads Πext\Pi_{\mathrm{ext}} of § 3.5, so that Π(q)=Πint(q)+Πcon(q)+Πext(q)\Pi(\mathbf{q}) = \Pi_{\mathrm{int}}(\mathbf{q}) + \Pi_{\mathrm{con}}(\mathbf{q}) + \Pi_{\mathrm{ext}}(\mathbf{q}). The structure is in equilibrium when this energy is stationary with respect to the DOF. We therefore define the residual

R(q)=∂Π∂q,(5)\mathcal{R}(\mathbf{q}) = \frac{\partial\Pi}{\partial\mathbf{q}}, \tag*{(5)}

and the equilibrium configuration q∗\mathbf{q}^* is the solution of R(q∗)=0\mathcal{R}(\mathbf{q}^*) = 0 subject to the prescribed boundary conditions. This system is nonlinear in q\mathbf{q} and is solved iteratively. At each iteration the tangent stiffness K(q)=∂2Π/∂q2\mathbf{K}(\mathbf{q}) = \partial^2\Pi/\partial\mathbf{q}^2 is assembled and the increment Δq=−K−1R\Delta\mathbf{q} = -\mathbf{K}^{-1}\mathcal{R} is applied, until the residual falls below a prescribed tolerance. Given the nonlinearity of the response, the actuation is applied incrementally over KK load steps, with the solution at each load step initialized from that of the preceding one. We denote the equilibrium configuration at load step kk as qk∗\mathbf{q}_k^*.

Actuation by Swelling

Having established the rod kinematics, we now describe the first of the two forms of actuation, in which programmability is achieved via the distribution of an active swellable gel and a passive elastomer matrix. Here, we briefly outline the formulation, with further details provided in our earlier work [25]. When a gel is immersed in a solvent bath, it absorbs solvent molecules and expands, while an elastomer remains largely passive (Figure 2). To quantify the swelling, we introduce a polymer volume fraction ϕ(s,ξ)∈(0,1]\phi(s,\xi) \in(0,1], where ϕ=1\phi= 1 is the fully dry state and ϕ→0\phi\to0 the fully swollen limit. Recall that molecular incompressibility prescribes the volume ratio in Equation (2) as J=1/ϕJ = 1/\phi.

Following Flory-Rehner theory [29, 30], the swelling contribution to the free energy density ψgel\psi_{\mathrm{gel}} comprises the energy of mixing the polymer with the solvent and the chemical work of exchanging solvent with the bath, and is given by

ψgel=RTΩ~[1−ϕϕln⁡(1−ϕ)+χ(1−ϕ)]−(μ−μ0)1Ω~(1ϕ−1),(6)\psi_{\mathrm{gel}} = \frac{RT}{\tilde{\Omega}}\left[\frac{1-\phi}{\phi} \ln(1-\phi) + \chi(1-\phi)\right] - (\mu-\mu^0)\frac{1}{\tilde{\Omega}}\left(\frac{1}{\phi}-1\right), \tag*{(6)}

where the solvent is characterized by its molar volume Ω~\tilde{\Omega} and reference chemical potential μ0\mu^0, the bath by its temperature TT and chemical potential μ\mu, and RR is the universal gas constant.

The material response is governed jointly by its Flory-Huggins interaction parameter χ\chi, which quantifies the affinity of the polymer network for the solvent, and the shear modulus GG associated with the elastic energy in Equation (3). A compliant gel network with low χ\chi readily takes up solvent and swells, whereas a stiff elastomer network with high χ\chi resists uptake and remains effectively passive. It is the contrast in GG and χ\chi between the two phases that produces shape change. Gel fibers take up solvent and expand while the surrounding elastomer resists. As the resulting differential expansion drives the deformation, the distribution of the two phases determines the deformed shape upon actuation.

Finally, the polymer volume fraction is determined by the thermodynamic equilibrium condition between the gel and the external solvent, namely the stationarity of the free energy density with respect to ϕ\phi. Applying ∂ψ/∂ϕ=0\partial\psi/\partial\phi= 0 to Equation (3) yields the implicit Flory-Rehner relation

F(ϕ):=μ0−μRT+ln⁡(1−ϕ)+ϕ+χϕ2+Ω~GRT(λt2−1)=0,λt2=1ϕλ.(7)\mathcal{F}(\phi) := \frac{\mu^0-\mu}{RT}+\ln(1-\phi)+\phi+\chi\phi^2+\frac{\widetilde{\Omega}G}{RT}\left(\lambda_t^2-1\right)=0,\qquad\lambda_t^2=\frac{1}{\phi}\lambda. \tag*{(7)}

Observe that ϕ\phi depends on the fiber stretch λ\lambda, so the swelling response is coupled to the mechanical state and is resolved together with the equilibrium configuration.

Magnetic Actuation

We now describe the second form of actuation, in which programmability is achieved via determining the distribution of oriented hard-magnetic particles within a soft matrix, together with the externally applied magnetic field. The following account builds on earlier treatments detailed in [31, 24, 1]. We consider ideal hard-magnetic soft materials, wherein the embedded particles are of high coercivity, with the permanent magnetization imprinted during fabrication. We further assume that the permeability of the composite is that of the ambient medium, and that magnetic interactions between distinct parts of the structure are negligible. We describe the magnetized state through the remnant magnetic flux density Br(s,ξ)B^r(s,\xi), specified in the global frame in the reference configuration and unaltered by the subsequent deformation of the structure. Actuation is in turn supplied by an applied magnetic flux density BaB^a, specified in the global frame as a sequence of spatially uniform fields, one per load step (Figure 2).

Following [24], the magnetic contribution ψmag\psi_{\mathrm{mag}} to Equation (3) is given by

ψmag=−1μ0ba⋅(Fbr),ba=DTBa,br=D0TBr,(8)\psi_{\mathrm{mag}}=-\frac{1}{\mu_0}b^a\cdot(Fb^r),\qquad b^a=D^TB^a,\qquad b^r=D_0^TB^r, \tag*{(8)}

where μ0\mu_0 is the vacuum permeability, and bab^a and brb^r are the applied and remnant flux densities resolved in the current and reference material frames respectively.

Since the applied field is spatially uniform, it exerts no net force on the structure. Instead, the actuation arises from the torques that the field exerts on the magnetized fibers, which tend to align the remnant magnetization with the applied field. Varying the remnant direction across the cross-section and along the length produces spatially varying torques under the applied field. The distribution of the remnant magnetization, together with the applied field, determines the actuated response.

Contacts and Applied Loadings

We now describe the interaction of the rod with its surroundings and with itself, together with the applied mechanical loading. Structures deployed within confined geometries or in the presence of obstacles can make contact with the surrounding boundary. Furthermore, structures undergoing large deformations may fold and make contact onto themselves. We consider both, referred to as external contact and self-contact, respectively. For external contact, we represent the surrounding geometry implicitly through a signed distance field φ(x)\varphi(x), adopting the convention that φ>0\varphi> 0 in free space. Furthermore, we assign to each element ee a contact radius rer_e of its cross-section.

The penetration depth of an element ee in external contact is ϵe=re−φ(xe)\epsilon_e=r_e-\varphi(x_e), where xex_e is the element center. For self-contact, the penetration depth of a pair of elements aa and bb is ϵab=(ra+rb)−hab\epsilon_{ab}=(r_a+r_b)-h_{ab}, where habh_{ab} is the minimum distance between the two element segments. In either case, the penetration is inactive when ϵ≤0\epsilon\leq0, and a penalty energy is accumulated otherwise. We adopt a C2C^2 continuous smooth ramp to regularize max⁡(ϵ,0)\max(\epsilon,0):

ρ(ϵ)={0ϵ≤0ϵ3w2−ϵ42w30<ϵ<wϵ−w2ϵ≥w(9)\rho(\epsilon)= \begin{cases} 0 & \epsilon\leq0\\ \frac{\epsilon^3}{w^2}-\frac{\epsilon^4}{2w^3} & 0<\epsilon<w\\ \epsilon-\frac{w}{2} & \epsilon\geq w \end{cases} \tag*{(9)}

where ww is a transition width taken as a fraction of the contact radius, and ρ\rho recovers max⁡(ϵ,0)\max(\epsilon,0) as w→0w\to0. The C2C^2 continuity ensures that the tangent stiffness remains continuous at the onset of contact and the sensitivities of § 3.8 are well defined.

The contact energy of an element, or of a pair of elements, is quadratic in the regularized penetration, and is given by πcon=12ωconℓcon2\pi_{\mathrm{con}}=\frac{1}{2}\omega_{\mathrm{con}}\ell_{\mathrm{con}}^2, where ωcon\omega_{\mathrm{con}} is a penalty stiffness and ℓcon\ell_{\mathrm{con}} is a representative contact length, taken as the element length for external contact and as the geometric mean of the two element lengths for self-contact. Summation over the elements and over the interacting element pairs gives Πcon(q)\Pi_{\mathrm{con}}(\boldsymbol{q}). To preserve the sparsity of the tangent stiffness, the self-contact contribution to the tangent is retained only in block-diagonal form. Furthermore, the contact force at a node ii then follows from the contact energy as

ficon=−∂Πcon∂δi.(10)\boldsymbol{f}^{\mathrm{con}}_i=-\frac{\partial\Pi_{\mathrm{con}}}{\partial\boldsymbol{\delta}^i}. \tag*{(10)}

where δi\boldsymbol{\delta}^i are the translational DOF of node ii. In other words, ficon\boldsymbol{f}^{\mathrm{con}}_i is the force exerted on the rod at node ii by the surroundings or by the rod itself, and is equal and opposite to the force that the rod exerts in return.

Finally, the applied mechanical loading contributes Πext(q)=−fTq\Pi_{\mathrm{ext}}(\boldsymbol{q})=-\boldsymbol{f}^{\mathrm{T}}\boldsymbol{q}, where f\boldsymbol{f} collects the prescribed nodal forces and moments. We treat the magnitude and direction of the applied loads as fixed as the structure deforms.

Optimization Formulation

We now integrate the preceding kinematic and actuation formulations to outline the proposed optimization framework. The numerical experiments in § 4 consider different design variables, objectives, and constraints; we present the complete set here and invoke the relevant subset in each study.

Design Variables

Recall that the design comprises the material assigned to every fiber segment, together with, for magnetic actuation, the applied magnetic field at every load step. We collect these into the vector of design variables ζ\boldsymbol{\zeta}. Since the underlying physical quantities differ in units and in magnitude, each design variable is normalized to [0,1][0,1]. In particular, a physical quantity pp with a prescribed extent [pmin⁡,pmax⁡][p_{\min},p_{\max}] is recovered from its design variable ζ\zeta as p=pmin⁡+ζ(pmax⁡−pmin⁡)p=p_{\min}+\zeta(p_{\max}-p_{\min}). This renders the design variables comparable in magnitude, and reduces the bounds on the physical quantities to the box constraint 0≤ζ≤10\leq\zeta\leq1.

For swelling actuation, a design variable is assigned to every fiber segment. The design variables are filtered and projected, as detailed in § 4, to obtain the pseudodensity ρef∈[0,1]\rho_{\mathrm{ef}}\in[0,1] of every fiber segment, where ρef=1\rho_{\mathrm{ef}}=1 corresponds to gel and ρef=0\rho_{\mathrm{ef}}=0 corresponds to elastomer. The shear modulus GefG_{\mathrm{ef}} and the Flory-Huggins parameter χef\chi_{\mathrm{ef}} are then interpolated between those of the two phases (Equation (20)).

For magnetic actuation, the design variables describe the remnant flux density Br\boldsymbol{B}^{\mathrm{r}} of the fiber segments, together with the applied flux density Bka\boldsymbol{B}^{\mathrm{a}}_k at every load step kk. Each flux density is parameterized by its magnitude, its azimuth α\alpha, and its polar angle β\beta (Figure 2). Here, the magnitude is bounded by a prescribed maximum, while the two angles span all unique orientations.

Objectives

Shape morphing: One of the central goals of the case studies showcased herein (§ 4.1 and § 4.2) is computing the optimal material distribution such that the structure morphs to a target shape upon actuation. Given a target centerline curvature κ^e\hat{\kappa}_e and the realized centerline curvature κe(q∗;ζ)\boldsymbol{\kappa}_e(\boldsymbol{q}^*;\zeta) at the element centers in the swollen configuration, the objective is expressed as the minimization of the relative squared error between the two fields:

J(ζ)=∑e=1Ne∥κe(q∗;ζ)−κ^e∥22∑e=1Ne∥κ^e∥22.(11)\mathcal{J}(\boldsymbol{\zeta})=\frac{\sum_{e=1}^{N_e}\left\|\boldsymbol{\kappa}_e(\boldsymbol{q}^*;\boldsymbol{\zeta})-\hat{\boldsymbol{\kappa}}_e\right\|_2^2}{\sum_{e=1}^{N_e}\left\|\hat{\boldsymbol{\kappa}}_e\right\|_2^2}. \tag*{(11)}

Contact forces: Another goal of the case studies showcased herein (§ 4.3) is computing the remnant magnetization and the applied magnetic field such that the forces exerted on the surroundings during the actuation are minimized. Recall that the contact force at a node is given by Equation (10). Denoting by fn,kcon\boldsymbol{f}^{\mathrm{con}}_{n,k} the contact force at node nn in the equilibrium configuration qk∗\boldsymbol{q}^*_k of load step kk, the objective is expressed as the minimization of the contact force summed over all the nodes and all the load steps:

J(ζ)=∑k=1K∑n=1Nn∥fn,kcon(qk∗;ζ)∥2.(12)\mathcal{J}(\boldsymbol{\zeta})=\sum_{k=1}^{K}\sum_{n=1}^{N_n}\left\|\boldsymbol{f}^{\mathrm{con}}_{n,k}(\boldsymbol{q}^*_k;\boldsymbol{\zeta})\right\|_2. \tag*{(12)}

Constraints

Grayness constraint: Recall that while the pseudodensities are relaxed to a continuous range (0≤ρef≤1)(0\leq\rho_{\mathrm{ef}\leq1}) to enable gradient-based optimization, physical realizability dictates that every fiber segment be composed of a single solid or liquid material. phase (ρef∈{0,1}\rho_{ef} \in\{0,1\}). To drive the design toward this binary state and eliminate intermediate pseudodensities, we introduce a grayness constraint [32]:

gr:=1NeNf∑e=1Ne∑f=1Nfρef(1−ρef)−ϵg∗≤0,(13)g_r := \frac{1}{N_eN_f}\sum_{e=1}^{N_e}\sum_{f=1}^{N_f}\rho_{ef}(1-\rho_{ef})-\epsilon_g^* \le0, \tag*{(13)}

where ϵg∗>0\epsilon_g^*>0 is the permitted grayness, governed by a continuation scheme. Initially, ϵg∗\epsilon_g^* is assigned a sufficiently large value to permit mixed material phases, enabling the optimizer to broadly explore the design space without prematurely trapping in local minima. Over the course of the optimization, ϵg∗\epsilon_g^* is systematically annealed toward zero. This progressive tightening imposes a strict penalty on gray regions, effectively resolving the design to a discrete partition.

Gel volume constraint: In some design scenarios, a constraint is imposed on the volume fraction of the gel, for instance to limit the usage of the active material. With AefA_{ef} and Δse\Delta s_e being the reference area and the reference length of a fiber segment, the constraint is expressed as:

gv:=∑e,fρefAefΔse∑e,fAefΔse−Vgel∗≤0,(14)g_v := \frac{\sum_{e,f}\rho_{ef}A_{ef}\Delta s_e}{\sum_{e,f}A_{ef}\Delta s_e}-V_{\mathrm{gel}}^* \le0, \tag*{(14)}

where Vgel∗∈(0,1]V_{\mathrm{gel}}^* \in(0,1] is the prescribed upper bound on the volume fraction of the gel.

Optimization Problem

Collecting the objectives (Equations (11) and (12)), the residual (Equation (5)), and the constraints (Equations (13) and (14)), the canonical optimization problem can be expressed as:

minimize⁡ζJ(ζ,{qk∗})(15)\underset{\zeta}{\operatorname{minimize}}\quad\mathcal{J}\left(\zeta,\{q_k^*\}\right) \tag*{(15)}
subject to⁡R(qk∗;ζ)=0,k=1,…,K(16)\operatorname{subject\ to}\quad\mathcal{R}(q_k^*;\zeta)=0,\quad k=1,\ldots,K \tag*{(16)}
gi(ζ)≤0,i∈{r,v}(17)g_i(\zeta)\le0,\quad i\in\{r,v\} \tag*{(17)}
0≤ζ≤1(18)0\le\zeta\le1 \tag*{(18)}

We employ the method of moving asymptotes (MMA) [33] to perform the design updates. The choices of the MMA move limit and the step tolerance are specified later under numerical experiments. Default values are used for other hyperparameters.

Sensitivity Analysis

We now detail the computation of the sensitivities of the objective and the constraints with respect to the design variables. Deriving and computing these is non-trivial in the present setting, owing to the geometric nonlinearity, the contact interactions, the coupling between deformation and swelling, and the sequence of equilibrium solves over which the response is evaluated. Towards this, we construct an end-to-end differentiable pipeline using the reverse-mode AD capabilities of JAX [34]. This eliminates the laborious and error-prone derivation of sensitivity expressions, and accommodates varying objectives, constraints, and discretizations.

Most of the computational pipeline, from the design vector through the material properties to the objective, is an explicit composition of differentiable operations to which the chain rule applies directly. However, two quantities do not admit a closed-form expression in the design variables. The first is the equilibrium configuration q∗q^*, which satisfies R(q∗;ζ)=0\mathcal{R}(q^*;\zeta)=0 and is resolved via Newton-Raphson iteration. The second is the polymer volume fraction ϕ∗\phi^*, which satisfies the Flory-Rehner relation of Equation (7) and is resolved via bisection. Both are iterative root-finding procedures. Differentiating through the iteration history is prohibitive in memory and compute. We instead apply the implicit function theorem (IFT) [35] uniformly to both.

For a converged solution y∗y^* satisfying H(y∗;ϑ)=0H(y^*;\vartheta)=0, and provided ∂H/∂y\partial H/\partial y is invertible at that state, the IFT gives

dy∗dϑ=−(∂H∂y)−1∂H∂ϑ,(19)\frac{\mathrm{d}y^*}{\mathrm{d}\vartheta}=-\left(\frac{\partial H}{\partial y}\right)^{-1}\frac{\partial H}{\partial\vartheta}, \tag*{(19)}

replacing differentiation through the iteration history with two partial derivatives evaluated at y⋆y^\star. Setting y=q⋆y=q^\star, H=R\mathcal{H}=\mathcal{R}, and ϑ=ζ\vartheta=\zeta, we obtain dq⋆/dζ=−K−1∂R/∂ζ\mathrm{d}q^\star/\mathrm{d}\zeta=-K^{-1}\partial\mathcal{R}/\partial\zeta, where K=∂R/∂qK=\partial\mathcal{R}/\partial q is the tangent stiffness at the converged configuration. Setting y=ϕ⋆y=\phi^\star, H=F\mathcal{H}=\mathcal{F}, and ϑ=(λ,G,χ)\vartheta=(\lambda,G,\chi), we obtain a scalar division at each fiber segment, since ∂F/∂ϕ\partial\mathcal{F}/\partial\phi is nonzero for all physically realizable states. The two rules are applied hierarchically. As ϕ\phi depends on the fiber stretch and hence on qq, its rule is invoked within the evaluation of ∂R/∂ζ\partial\mathcal{R}/\partial\zeta.

Finally, recall that the actuation is applied as a sequence of load steps. The sensitivity propagates through the corresponding sequence of equilibrium solves in reverse, at a cost of one linear solve per load step.

Algorithm

We assume a prescribed reference configuration of the rod, the candidate materials, the boundary conditions, the actuation, and, where relevant, the geometry of the surroundings. The procedure is as follows:

  1. Rod initialization: The reference centerline and directors are defined, and the centerline is discretized into NeN_e elements (§ 3.1). The cross-section is resolved into NfN_f fibers with prescribed offsets and areas. The fixtures and the applied loads are prescribed.

  1. Material and actuation initialization: The material constants are specified. For swelling actuation, these include the shear moduli and the Flory-Huggins parameters of the gel and elastomer phases, together with the molar volume of the solvent, the temperature, and the chemical potentials of the dry and wet states. For magnetic actuation, these include the shear modulus and the bounds on the remnant and applied flux densities. Where relevant, the signed distance field of the surroundings and the contact parameters are specified (§ 3.5).

  1. Solvers and optimizer: The tolerances for the bisection solver and the Newton-Raphson iteration are specified, along with the number of load steps KK. The optimization settings are then prescribed, including the filter and projection parameters, the constraint bounds, and the MMA move limit. Note that we adopt continuation schemes for the projection and the grayness constraint.

  1. Design variable initialization: The normalized design variables ζ\zeta are initialized (§ 3.6.1).

  1. Property mapping: At each iteration, the design variables are mapped to the physical quantities. For swelling actuation, the design variables are filtered and projected to obtain the pseudodensities ρef\rho_{ef}, from which the shear modulus GefG_{ef} and the Flory-Huggins parameter χef\chi_{ef} of every fiber segment are interpolated (Equation (20)). For magnetic actuation, the design variables are renormalized to obtain the remnant flux density of every fiber segment and the applied flux density at every load step.

  1. Equilibrium analysis: The actuation is applied incrementally over the KK load steps. At each load step, the equilibrium configuration qk⋆q_k^\star is obtained by solving Equation (5) via the Newton-Raphson method, starting from the configuration of the preceding load step. Within each Newton-Raphson iteration, the strains are projected onto the fiber segments to assemble the deformation gradient (Equation (1)). Where swelling is present, the Flory-Rehner relation (Equation (7)) is solved for the polymer volume fraction. The free energy density (Equation (3)) is then evaluated, with the swelling and magnetic contributions given by Equation (6) and Equation (8). Finally, the internal energy (Equation (4)) is assembled together with the contact energy (§ 3.5) and the work of the applied loads, from which the residual and the tangent stiffness follow.

  1. Objective and constraints: The relevant objective (Equation (11) or Equation (12)) is evaluated from the equilibrium configurations, and the relevant constraints (Equations (13) and (14)) are evaluated from the pseudodensities.

  1. Sensitivity analysis: The sensitivities of the objective and the constraints with respect to the design variables are computed via reverse-mode AD through our end-to-end differentiable pipeline, as described in § 3.8.

  1. Design update: The objective, the constraints, and their sensitivities are passed to MMA, which computes the updated design variables. The parameters governed by a continuation scheme are then advanced.

Steps 5 through 9 are repeated until the convergence criterion is satisfied or the maximum number of iterations is reached.

Numerical Studies

In this section, we present three case studies to illustrate the proposed framework. All case studies are conducted on an AMD EPYC 7413 CPU using the JAX library [34] in Python. The three case studies are summarized in Table 1, and are ordered by increasing complexity.

Case I (Section 3.1)Case II (Section 3.2)Case III (Section 3.3)
Application“NU” shape morphingPatient-specific stentsMagnetically steered catheters
DemonstratesInterpretable shape morphingPatient-specific shape morphingMaterial-stimulus co-design
DeformationPlanar (xx-zz)3D3D
Cross-sectionRectangularAnnular latticeAnnular
Aspect ratio A\mathcal{A}L/H=200,250,300L/H = 200, 250, 300L/R=35L/R = 35L/Ro=175L/R_{\mathrm{o}} = 175 to 270270
Ne×NfN_{\mathrm{e}} \times N_{\mathrm{f}}300×20300 \times20300×180300 \times180120×8120 \times8
ActuationSwellingSwellingMagnetic field
Contact--Rigid, frictionless wall
Design variablesρef\rho_{\mathrm{ef}} per fiber segmentρef\rho_{\mathrm{ef}} per fiber segment on latticeBerB_{\mathrm{e}}^{r} per element, BkaB_{\mathrm{k}}^{a} per increment
ObjectiveCurvature mismatch (eq. (11))Curvature mismatch (eq. (11))Wall force (eq. (12))

Table 1. Summary of the case studies.

Furthermore, the default parameters shared across the case studies are as follows, unless otherwise specified:

  • Reference configuration: The rod is straight in its stress-free undeformed reference configuration.

  • Design variables: The design variables ζ\zeta are normalized to [0,1][0,1] and initialized with 0.5.

  • Filter: To ensure smoothness, the design variables are filtered along the rod axis using a linear density filter with a radius of 2L/Ne2L/N_{\mathrm{e}}.

  • Projection: For the swelling experiments, the filtered variables are thresholded using a smoothed Heaviside projection [36] with the sharpness increased linearly from 1 to 128 over the first 60 iterations.

  • Materials: All materials are modeled as neo-Hookean with molecular incompressibility. For the swelling experiments (§ 4.1 and § 4.2), the gel and elastomer Flory-Huggins parameters are set to χgel=0.2\chi_{\mathrm{gel}} = 0.2 and χelast=2.5\chi_{\mathrm{elast}} = 2.5, and the shear modulus as Ggel=1G_{\mathrm{gel}} = 1 MPa and Gelast=50G_{\mathrm{elast}} = 50 MPa, respectively. For the magnetic experiment (§ 4.3), the shear modulus is set to G=303G = 303 kPa [1].

  • Solvent: For swelling, a solvent with molar volume Ω~=1.8×10−5 m3 mol−1\widetilde{\Omega} = 1.8 \times10^{-5}\ \mathrm{m}^{3}\ \mathrm{mol}^{-1} at a temperature T=310T = 310 K is assumed, with μ0=0\mu^{0} = 0, μdry=−1×105 J mol−1\mu_{\mathrm{dry}} = -1 \times10^{5}\ \mathrm{J}\ \mathrm{mol}^{-1}, and μwet=−100 J mol−1\mu_{\mathrm{wet}} = -100\ \mathrm{J}\ \mathrm{mol}^{-1} (Equation (7)). The chemical potential μ\mu is increased geometrically from μdry\mu_{\mathrm{dry}} to μwet\mu_{\mathrm{wet}} over 70 load steps.

  • Nonlinear solver: The equilibrium equations are solved using the Newton-Raphson method with Armijo backtracking to a relative residual tolerance of 10−810^{-8}. A direct sparse solver is enlisted for the linear solves.

  • Grayness constraint: For the swelling experiments, the grayness slack ϵg∗\epsilon_{\mathrm{g}}^{*} is decreased linearly from 0.3 to 0.02 over the first 60 iterations.

  • Optimizer: The optimization is performed using MMA [33], and is terminated when the KKT residual or the change in the design variables falls below 10−610^{-6}, or when the maximum number of 300 iterations is reached.

Case I: Planar Shape Morphing of a Layered Rod

Recall that the central hypothesis of our work is that Cosserat rods are well suited for the representation, design, and optimization of slender structures. While conventional continuum FEA requires meshing the cross-section, the proposed formulation discretizes only the centerline, accounting for the fibers solely through integration over the cross-section (§ 3.1). To illustrate this, we begin by considering a simple planar example, where a slender rod morphs to a target shape upon swelling (Table 1). In this case study, the target shape upon deformation is chosen such that the optimized material layout can be interpreted by inspection.

Consider a straight rod of length L=138L = 138 mm with a rectangular cross-section of width WW and height HH, clamped at its left end, as illustrated in Figure 5(A). For illustration, we consider three slender aspect ratios, A:=L/H=300\mathcal{A} := L/H = 300, 250, and 200, obtained by varying HH with LL fixed and W=0.7HW = 0.7H. The rod centerline is discretized into Ne=300N_{\mathrm{e}} = 300 elements, and the cross-section is discretized into Nf=20N_{\mathrm{f}} = 20 fibers of uniform height H/NfH/N_{\mathrm{f}} stacked along the height. The rod swells freely and is restricted to deform in the xx-zz plane. The target shape “NU” is illustrated in Figure 5(B).

Planar shape morphing of a slender rod

Figure 5. Planar shape morphing of a slender rod. (A) The rod and its cross-section. (B) The target shape. (C) Swollen configurations of the optimized designs for three aspect ratios, A=300\mathcal{A}=300, 250, 200. (D)–(G) show further analyses for A=200\mathcal{A}=200. (D) Optimized material layout in the undeformed configuration. (E) Realized and target curvature. (F) Convergence of the objective and the constraints. (G) Designs at select iterations.

To program the target deformation, we assign a design variable to every fiber of every element, resulting in 6000 design variables ζ\zeta, while the degrees of freedom remain those of the centerline. The design variables are filtered along the length of the rod and projected to obtain the pseudodensity ρef\rho_{\mathrm{ef}} of fiber ff in element ee, where ρef=1\rho_{\mathrm{ef}} = 1 corresponds to gel and ρef=0\rho_{\mathrm{ef}} = 0 corresponds to elastomer. The shear modulus GG and the Flory-Huggins parameter χ\chi of each fiber segment are then interpolated as

Equation not extracted
Gef=ρefGgel+(1−ρef)Gelast,χef=ρefχgel+(1−ρef)χelast.(20)G_{ef}=\rho_{ef}G_{gel}+(1-\rho_{ef})G_{elast},\qquad\chi_{ef}=\rho_{ef}\chi_{gel}+(1-\rho_{ef})\chi_{elast}. \tag*{(20)}

We adopt the shape-morphing objective in Equation (11), where the mismatch between the realized and target curvature is minimized. In addition to the grayness constraint, the gel volume fraction is bounded from above to a nominal value of 0.4. The optimization is performed with an MMA move limit of 5×10−45\times10^{-4}.

The swollen configurations of the optimized designs for the three aspect ratios are illustrated in Figure 5(C), overlaid on the target shape. In all three cases, we confirm strong match between the target shape and the optimized result. For brevity, the remaining analyses in Figure 5 are only presented for A=200\mathcal{A}=200. The optimized material layout in the undeformed configuration is illustrated in Figure 5(D). The curvature distributions are compared in Figure 5(E), which shows agreement along the entire length of the rod. As expected, the gel is placed on the convex side of each bend, while the straight segments remain elastomer. The convergence of the objective and the constraints is illustrated in Figure 5(F). The optimization reduces the objective by several orders of magnitude over 250 iterations, with a final grayness of 0.02 and a gel volume fraction of 0.14. Note that the gel constraint remains inactive. Similar convergence characteristics were observed for the other aspect ratios. Finally, the evolution of the material distribution at select iterations during the optimization is illustrated in Figure 5(G), where the initially uniform straight rod progressively morphs towards the target. Having illustrated the framework on a planar example, we now turn to the three-dimensional design of a shape-morphing stent.

Case II: Conforming Swelling-Programmed Stents

As a second case study, we consider the design of patient-specific shape-morphing stents. A stent is a mesh tube that is inserted into a vessel to counteract a disease-induced constriction of the flow [37]. Once deployed, a stent with low bending flexibility can permanently straighten the vessel and injure its wall [38, 39]. A direct remedy is a stent whose unloaded shape upon deployment is that of the vessel. Since the shape of the vessel differs from patient to patient, such a stent must be personalized. The personalization of stents and the use of the material distribution as a design variable have been identified as open directions in stent design [38]. Motivated by both the line of prior work and recent demonstrations of hydrogel-based shape morphing stents [40, 41], we seek a gel-elastomer distribution such that a stent that is straight when dry morphs to the shape of the patient’s vessel upon swelling in body fluid.

The stent considered is illustrated in Figure 6. We adopt the lattice of a commercial Express-type stent [19], where the unit cell is repeated 12 times along the length and 6 times along the circumference. Note that the lattice is kept fixed, and only the material distribution is optimized. The stent has an assumed aspect ratio A:=L/R=35\mathcal{A} := L/R = 35, where LL and RR denote its length and radius. Considering only the deformation of the centerline of this slender stent, we model the stent as a single Cosserat rod with an annular cross-section. The rod is discretized into Ne=300N_e = 300 elements, and the cross-section is discretized into Nf=180N_f = 180 fibers along the circumference, where the fiber segments that do not lie on the lattice are treated as void.

The stent considered in the dry state: side view, cross-section, and unit cell of the Express-type lattice

Figure 6. The stent considered in the dry state. (Left) Side view. (Middle) Cross-section. (Right) Unit cell of the Express-type lattice [19].

The target three-dimensional shapes are derived from the carotid dataset in [42], where the lumen and the centerlines of the vessels were reconstructed from computed tomography angiography scans of patients. Herein, we present the results on four selected tortuous segments, each from a different patient, as highlighted in Figure 7, and extract the target curvature from their centerlines. The length LL of each stent is set to the length of the corresponding segment, and its radius is specified by A\mathcal{A}. Furthermore, the target is assumed to be free of twist.

Patient-specific shape-morphing stents

Figure 7. Patient-specific shape-morphing stents, where rows A to D correspond to four patients. For each patient: (left) the carotid vessel with the selected segment highlighted; (middle, top) the realized and target curvatures; (middle, bottom) the optimized material layout unrolled along the circumference; (right) four views of the swollen stent.

As in § 4.1, we assign a design variable to every fiber segment on the lattice, and adopt the material interpolation in Equation (20) and the shape-morphing objective in Equation (11) (Table 1). Here, the design variables are filtered along the length with a radius of 0.5 mm, and the mismatch between the realized and target curvatures is minimized over the two bending components and the twist. Furthermore, only the grayness constraint is imposed, and the optimization is performed with an MMA move limit of 10−310^{-3}.

The optimized designs for the four patients are illustrated in rows A to D of Figure 7. For each patient, we illustrate the vessel with the selected segment, the realized and target curvatures, the optimized material layout unrolled along the circumference, and four views of the swollen stent. Observe that the realized bending curvatures κ1\kappa_1 and κ2\kappa_2 closely follow the target, with some local discrepancies. In contrast to § 4.1, the optimized layouts cannot be readily interpreted by inspection, which underscores the need for optimization. Furthermore, observe that a distinct layout is obtained for each patient on the same lattice, which highlights the patient-specific customization enabled by the framework. Convergence characteristics similar to those in § 4.1 were observed in all the cases.

We remark that the present study only targets the conformity of the stent to the shape of the vessel. Other considerations pertaining to the deployment of the stent, its post-deployment structural response such as radial strength and fatigue, and the hemodynamics [38] are not considered in the present scope, and could be future work.

Case III: Magnetically Steered Catheters for Reduced Wall Contact

As a final case study (Table 1), we consider the design of magnetically steered catheters for minimally invasive procedures. In endovascular procedures, a catheter is advanced through tortuous vessels to reach a target site, where it is steered primarily through its contact with the vessel wall [4]. However, the contact between the catheter and the vessel wall can injure the wall, and has been associated with complications such as dissection, perforation, thrombosis, and embolic stroke [43]. Catheters with hard-magnetic particles embedded in a soft body can instead be steered remotely through an applied magnetic field [4, 13, 22]. Here, we seek to design the remnant magnetization of the catheter and the applied field such that the normal contact force exerted on the vessel wall during the insertion is minimized.

The vessels considered are illustrated in Figure 8. As in § 4.2, the vessels are derived from the dataset in [42]. For each of the four patients, we extract a path along the centerline, and model the vessel as a tube with a circular lumen of radius 1.51.5 mm along this path. The vessel wall is assumed to be rigid and frictionless. The catheter is assumed to be a slender annular tube with an outer radius Ro=0.5R_o = 0.5 mm and an inner radius Ri=0.35R_i = 0.35 mm. The aspect ratio A:=L/Ro\mathcal{A} := L/R_o ranges from 175 to 270 in the cases considered, which makes it well suited to the proposed formulation targeting slender structures. The catheter is discretized into Ne=120N_e = 120 elements, and its cross-section is discretized into Nf=8N_f = 8 fibers along the circumference. The catheter is advanced quasi-statically into the vessel along the tangent at the entry in K=382K = 382 increments. The contact with the vessel wall is modeled as described in § 3.5, where the contact radius is set to re=Ror_e = R_o, the transition width to w=0.25rew = 0.25r_e, and the penalty stiffness to ωcon=EI/ℓcon4\omega_{\mathrm{con}} = EI/\ell_{\mathrm{con}}^4, with E=3GE = 3G and II the second moment of area of the cross-section.

Magnetically steered catheters for four patients

Figure 8. Magnetically steered catheters, where rows A to D correspond to four patients. For each patient: (left) the carotid vessel with the selected path highlighted, and the idealized vessel considered in the simulation; (middle, top) the wall force as a function of the insertion depth; (middle, bottom) the optimized remnant flux density along the catheter and the optimized applied flux density over the insertion, resolved along the tangent and the two directors; (right) the contact forces of the unmagnetized and the optimized catheters at 25%, 50%, 75%, and 100% of the maximum insertion depth, where the size and the color of each sphere denote the contact force.

The design variables ζ\boldsymbol{\zeta} entail both the remnant magnetization of the catheter and the applied external magnetic field, forming a material-stimulus co-design problem [1, 10]. In particular, we assign a remnant flux density Ber\mathbf{B}^{\mathrm{r}}_e to every element ee, with an illustrative ∥Ber∥≤200\lVert\mathbf{B}^{\mathrm{r}}_e\rVert\le200 mT. Here, Ber\mathbf{B}^{\mathrm{r}}_e is specified in the global frame in the reference configuration and is uniform over the cross-section. We also assign a spatially uniform applied flux density Bka\mathbf{B}^{\mathrm{a}}_k to every increment k=1,…,Kk = 1,\ldots,K, with an illustrative ∥Bka∥≤20\lVert\mathbf{B}^{\mathrm{a}}_k\rVert\le20 mT. Each flux density is parameterized by its magnitude, polar angle, and azimuthal angle, resulting in 3Ne+3K3N_e + 3K design variables. The remnant and applied flux densities are filtered with a radius of 5 mm along the length of the catheter and along the insertion depth, respectively. We adopt the objective in Equation (12), where the wall force summed over all the increments is minimized. No constraints are imposed. The optimization is performed with an MMA move limit of 10−210^{-2}.

The optimized designs for the four vessel segments from different patients are illustrated in rows A to D of Figure 8. For each vessel segment, we illustrate the vessel with the prescribed path, the wall force as a function of the insertion depth dd, the optimized remnant and applied flux densities, and the contact forces at four insertion depths. Here, an unmagnetized catheter, which is steered by the contact with the wall alone, serves as the reference. Observe that the optimized designs reduce the wall force over most of the insertion. In particular, the objective is reduced by 33%, 50%, 43%, and 63% in cases A to D, respectively. The contact forces are illustrated at 25%, 50%, 75%, and 100% of the maximum insertion depth, where the size and the color of each sphere denote the contact force at that location. For instance, in case A at d=78d = 78 mm, the wall force reduces from 14.09 mN to 6.24 mN. Observe also that the contact is reduced along the body of the catheter, and not only at its tip.

We remark that the objective is the wall force summed over the insertion, which does not guarantee that the wall force is reduced at every depth, a concern since perforation occurs once the local contact force exceeds a threshold [44]. The minimization of the peak wall force is left for future work. Furthermore, other considerations such as the friction with the vessel wall and the deformation of the vessel wall [45] are left as future work.

Conclusion

This work proposes CROPS, a Cosserat-rods-based framework for the representation, design, and optimization of programmable slender structures. The representation captures large translations and rotations of slender structures while discretizing only the centerline. Furthermore, the cross-section of the rod is resolved into fibers, allowing the material to vary both along the length and across the cross-section. Two forms of actuation are considered, namely the swelling of gel-elastomer composites and the magnetic actuation of hard-magnetic composites. The material layout, together with the applied actuation, programs the response of the structure. Additionally, the contact of the rod with itself and its surroundings is accounted for. Finally, the framework is implemented in an end-to-end differentiable manner, with the sensitivities computed via automatic differentiation, enabling the gradient-based optimization of the material layout and the actuation. We demonstrate the framework for patient-specific programmable devices. First, stents are programmed to morph to the shape of carotid vessels upon swelling. Second, the remnant magnetization and the applied field of magnetically steered catheters are co-designed to reduce the force exerted on the vessel wall during insertion.

Several directions remain open to extend this further. The fidelity of the contact model can be improved by incorporating friction, deformable contact bodies, and the interaction with a surrounding fluid [11]. Furthermore, the design formulation can be extended. For instance, the peak wall force can be minimized in place of the summed wall force, and stent requirements such as radial strength and fatigue [38] can be imposed alongside the target shape. Similarly, applied fields with spatial gradients and closed-loop control of the applied field [46] would offer finer control over the catheter. Additionally, extending the framework to assemblies of multiple rods and to dynamics [47] would broaden its scope to untethered soft robots [48]. Finally, the fabrication and experimental testing of the optimized designs [40, 4] is essential to validate the framework. Ultimately, CROPS provides a differentiable platform for the design of programmable slender structures, with direct applicability to minimally invasive devices, soft robotics, and deployable structures.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Compliance with ethical standards

The authors declare that they have no conflict of interest.

Acknowledgments

A.C. and W.C. acknowledge the support of DARPA METALS under agreement No. HR0011-24-2-0302. W.C., D.D.L., S.L. and H.K. also acknowledge support from the NSF Boosting Research Ideas for Transformative and Equitable Advances in Engineering (BRITE) Fellow Program (CMMI-2227641). W.C. and D.D.L. are also grateful for support from NASA’s Minority University Research and Education Project Institutional Research Opportunity (MIRO) through the Center for In-Space Manufacturing (CISM-R2): Recycling and Regolith Processing (Award 80NSSC24M0176). W.C. acknowledges support from the NSF Materials Research Science and Engineering Centers (MRSEC) program (DMR-2308691).

Declaration of generative AI and AI-assisted technologies

During the preparation of this manuscript, the authors used Google’s Gemini and Anthropic’s Claude to improve the language and readability. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

Replication of Results

The Python source code is available at github.com/ideal-nu/crops.

References

References

  1. [1]Liwei Wang. Co-design of magnetic soft robots with large deformation and contacts via material point method and topology optimization. Computer Methods in Applied Mechanics and Engineering, 445:118205, 2025.DOI
  2. [2]Yue Dong, Lu Wang, Neng Xia, Zhengxin Yang, Chong Zhang, Chengfeng Pan, Dongdong Jin, Jiachen Zhang, Carmel Majidi, and Li Zhang. Untethered small-scale magnetic soft robot with programmable magnetization and integrated multifunctional modules. Science advances, 8(25):eabn8932, 2022.DOI
  3. [3]Wenqi Hu, Guo Zhan Lum, Massimo Mastrangeli, and Metin Sitti. Small-scale soft-bodied robot with multimodal locomotion. Nature, 554(7690):81–85, 2018.DOI
  4. [4]Yoonho Kim, German A Parada, Shengduo Liu, and Xuanhe Zhao. Ferromagnetic soft continuum robots. Science robotics, 4(33):eaax7329, 2019.DOI
  5. [5]Jay Sim, Shuai Wu, Sarah Hwang, Lu Lu, and Ruike Renee Zhao. Selective actuation enabled multifunctional magneto-mechanical metamaterial for programming elastic wave propagation. Advanced Functional Materials, 35(16):2422325, 2025.DOI
  6. [6]Zhi Zhao and Xiaojia Shelly Zhang. Encoding reprogrammable properties into magneto-mechanical materials via topology optimization. npj Computational Materials, 9(1):57, 2023.DOI
  7. [7]J William Boley, Wim M Van Rees, Charles Lissandrello, Mark N Horenstein, Ryan L Truby, Arda Kotikian, Jennifer A Lewis, and L Mahadevan. Shape-shifting structured lattices via multimaterial 4d printing. Proceedings of the National Academy of Sciences, 116(42):20856–20862, 2019.DOI
  8. [8]A Sydney Gladman, Elisabetta A Matsumoto, Ralph G Nuzzo, Lakshminarayanan Mahadevan, and Jennifer A Lewis. Biomimetic 4d printing. Nature materials, 15(4):413–418, 2016.DOI
  9. [9]Yoonho Kim, Hyunwoo Yuk, Ruike Zhao, Shawn A Chester, and Xuanhe Zhao. Printing ferromagnetic domains for untethered fast-transforming soft materials. Nature, 558(7709):274–279, 2018.DOI
  10. [10]Dex Doksoo Lee, Liwei Wang, Bolin Chen, Ying-Kuan Tsai, Jay Sim, Ruike Renee Zhao, Xiaojia Shelly Zhang, and Wei Chen. Co-design of programmable material systems. npj Metamaterials, 2(1):29, 2026.DOI
  11. [11]Mattia Gazzola, Levi H Dudte, Andrew G McCormick, and Lakshminarayanan Mahadevan. Forward and inverse problems in the mechanics of soft filaments. Royal Society open science, 5(6):171628, 2018.DOI
  12. [12]Oliver Weeger, Yue Sheng Benjamin Kang, Sai-Kit Yeung, and Martin L Dunn. Optimal design and manufacture of active rod structures with spatially variable materials. 3D Printing and Additive Manufacturing, 3(4):204–215, 2016.DOI
  13. [13]Liu Wang, Dongchang Zheng, Pablo Harker, Aman B Patel, Chuan Fei Guo, and Xuanhe Zhao. Evolutionary design of magnetic soft continuum robots. Proceedings of the National Academy of Sciences, 118(21):e2021922118, 2021.DOI
  14. [14]Artinian Azad, Huet Quentin, Ben Amar Faiz, and Perdereau Véronique. Optimal cosserat-based deformation control for robotic manipulation of linear objects. In 2023 IEEE/ASME International Conference on Advanced Intelligent Mechatronics (AIM), pages 381–388. IEEE, 2023.DOI
  15. [15]Juan C Alzate Cobo, Xiang-Long Peng, Bai-Xiang Xu, and Oliver Weeger. A finite swelling 3d beam model with axial and radial diffusion. Computer Methods in Applied Mechanics and Engineering, 441:117983, 2025.
  16. [16]Lingxiao Xun, Gang Zheng, and Alexandre Kruszewski. Cosserat-rod-based dynamic modeling of soft slender robot interacting with environment. IEEE Transactions on Robotics, 40:2811–2830, 2024.arxiv.org/abs/2307.06261
  17. [17]HE Goicoechea, Fernando Salvador Buezas, and Marta Beatriz Rosales. A non-linear cosserat rod model for drill-string dynamics in arbitrary borehole geometries with contact and friction. International Journal of Mechanical Sciences, 157:98–110, 2019.
  18. [18]Sunčica Čanić, Luka Grubišić, Domagoj Lacmanović, Matko Ljulj, and Josip Tambača. Optimal design of vascular stents using a network of 1d slender curved rods. Computer methods in applied mechanics and engineering, 394:114853, 2022.
  19. [19]Sunčica Čanić, Luka Grubišić, Matko Ljulj, Marcel Maretić, and Josip Tambača. Geometric optimization of vascular stents modeled as networks of 1d rods. Journal of computational physics, 494:112497, 2023.
  20. [20]Dario Carbonaro, Francesco Mezzadri, Nicola Ferro, Giuseppe De Nisco, Alberto Luigi Audenino, Diego Gallo, Claudio Chiastra, Umberto Morbiducci, and Simona Perotto. Design of innovative self-expandable femoral stents using inverse homogenization topology optimization. Computer Methods in Applied Mechanics and Engineering, 416:116288, 2023.
  21. [21]Yuhang Liu, Jiaheng Wang, Xiwen Lu, Yunlong Zhu, and Chenyao Bai. A unified optimization strategy for the geometry and magnetization of magnetic soft continuum robots in vascular intervention. Iscience, 28(5), 2025.
  22. [22]Giovanni Pittiglio, Peter Lloyd, Tomas da Veiga, Onaizah Onaizah, Cecilia Pompili, James H Chandler, and Pietro Valdastri. Patient-specific magnetic catheters for atraumatic autonomous endoscopy. Soft robotics, 9(6):1120–1133, 2022.
  23. [23]Joshua Davy, Peter Lloyd, James H Chandler, and Pietro Valdastri. Computational design of magnetic soft shape-forming catheters using the material point method. arXiv preprint arXiv:2310.19983, 2023.arxiv.org/abs/2310.19983
  24. [24]Zhi Zhao and Xiaojia Shelly Zhang. Topology optimization of hard-magnetic soft materials. Journal of the Mechanics and Physics of Solids, 158:104628, 2022.DOI
  25. [25]Aaditya Chandrasekhar, Dex Doksoo Lee, Hyunwoo Kwon, and Wei Chen. GELATO: Multi-material topology optimization of programmable gel-elastomer structures. Structural and Multidisciplinary Optimization, 69(9):214, September 2026.
  26. [26]Oliver Weeger. Isogeometric sizing and shape optimization of 3d beams and lattice structures at large deformations. Structural and Multidisciplinary Optimization, 65(2):43, 2022.
  27. [27]Myung-Jin Choi and Seonho Cho. Isogeometric configuration design sensitivity analysis of geometrically exact shear-deformable beam structures. Computer Methods in Applied Mechanics and Engineering, 351:153–183, 2019.
  28. [28]Juan C Simo and Loc Vu-Quoc. A three-dimensional finite-strain rod model. part ii: Computational aspects. Computer methods in applied mechanics and engineering, 58(1):79–116, 1986.
  29. [29]Paul J Flory and John Rehner Jr. Statistical mechanics of cross-linked polymer networks ii. swelling. The journal of chemical physics, 11(11):521–526, 1943.DOI
  30. [30]Wei Hong, Xuanhe Zhao, Jinxiong Zhou, and Zhigang Suo. A theory of coupled diffusion and large deformation in polymeric gels. Journal of the Mechanics and Physics of Solids, 56(5):1779–1793, May 2008.
  31. [31]Ruike Zhao, Yoonho Kim, Shawn A Chester, Pradeep Sharma, and Xuanhe Zhao. Mechanics of hard-magnetic soft materials. Journal of the Mechanics and Physics of Solids, 124:244–263, 2019.
  32. [32]Xiaoping Qian. Undercut and overhang angle control in topology optimization: a density gradient based integral approach. International Journal for Numerical Methods in Engineering, 111(3):247–272, 2017.DOI
  33. [33]Krister Svanberg. The method of moving asymptotes—a new method for structural optimization. International journal for numerical methods in engineering, 24(2):359–373, 1987.DOI
  34. [34]James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018.
  35. [35]Mathieu Blondel, Quentin Berthet, Marco Cuturi, Roy Frostig, Stephan Hoyer, Felipe Llinares-López, Fabian Pedregosa, and Jean-Philippe Vert. Efficient and modular implicit differentiation. Advances in neural information processing systems, 35:5230–5242, 2022.arxiv.org/abs/2105.15183
  36. [36]Fengwen Wang, Boyan Stefanov Lazarov, and Ole Sigmund. On projection methods, convergence and robust formulations in topology optimization. Structural and multidisciplinary optimization, 43(6):767–784, 2011.DOI
  37. [37]Josip Tambača, Mate Kosor, Sunčica Čanić, and D Paniagua MD. Mathematical modeling of vascular stents. SIAM journal on applied mathematics, 70(6):1922–1952, 2010.DOI
  38. [38]Ankush Kapoor, Nigel Jepson, Neil W Bressloff, Poay Huan Loh, Tapabrata Ray, and Susann Beier. The road to the ideal stent: A review of stent design optimisation methods, findings, and opportunities. Materials & Design, 237:112556, 2024.DOI
  39. [39]Ferdinando Auricchio, Michele Conti, Anna Ferrara, Simone Morganti, and Alessandro Reali. Patient-specific finite element analysis of carotid artery stenting: a focus on vessel modeling. International Journal for Numerical Methods in Biomedical Engineering, 29(6):645–664, 2013.DOI
  40. [40]Bo Liu, Hui Li, Fengzhen Meng, Ziyang Xu, Liuzhi Hao, Yuan Yao, Hao Zhu, Chenmin Wang, Jun Wu, Shaoquan Bian, William W. Lu, Wenguang Liu, Haobo Pan, and Xiaoli Zhao. 4D printed hydrogel scaffold with swelling-stiffening properties and programmable deformation for minimally invasive implantation. Nature Communications, 15(1):1587, February 2024.DOI
  41. [41]Yoshiyasu Nagakawa, Satoshi Fujita, Shunji Yunoki, Takayoshi Tsuchiya, Shin-ichiro Suye, Kenji Kinoshita, Motoki Sasaki, and Takao Itoi. Characterization and preliminary in vivo evaluation of a self-expandable hydrogel stent with anisotropic swelling behavior and endoscopic deliverability for use in biliary drainage. Journal of Materials Chemistry B, 10(23):4375–4385, June 2022.DOI
  42. [42]Pepe Eulzer, Kevin Richter, Tristan Probst, Anna Hundertmark, and Kai Lawonn. A dataset of reconstructed carotid bifurcation lumen and plaque models with centerline tree and simulated hemodynamics, February 2024.DOI
  43. [43]Masoud Razban, Javad Dargahi, and Benoit Boulet. Image-based intraluminal contact force monitoring in robotic vascular navigation. arXiv preprint arXiv:2012.10762, 2020.arxiv.org/abs/2012.10762
  44. [44]Francesco Perna, E Kevin Heist, Stephan B Danik, Conor D Barrett, Jeremy N Ruskin, and Moussa Mansour. Assessment of catheter tip contact force resulting in cardiac perforation in swine atria using force sensing technology. Circulation. Arrhythmia and electrophysiology, 4(2), 2011.DOI
  45. [45]Chengxiong Lin, Huilu Yuan, and Chengyong Wang. The study of quantitative friction damage models at the interventional catheter–vascular tissue interface. Friction, 14(8):9441194, November 2025.DOI
  46. [46]Vittorio Francescon, Nikita Murasovs, Peter Lloyd, Onaizah Onaizah, Damith Suresh Chathuranga, and Pietro Valdastri. Closed-loop shape-forming control of a magnetic soft continuum robot. IEEE Robotics and Automation Letters, 10(6):6071–6078, 2025.
  47. [47]Tejaswin Parthasarathy, Seung Hyun Kim, Songyuan Cui, and Mattia Gazzola. Elastica++: A high-performance, multiphysics framework for large interacting assemblies of cosserat rods. arXiv preprint arXiv:2605.13766, 2026.arxiv.org/abs/2605.13766
  48. [48]Taekyoung Kim, Pranav Kaarthik, and Ryan L Truby. A flexible, architected soft robotic actuator for motorized extensional motion. Advanced Intelligent Systems, 6(11):2300866, 2024.

Paper details

vlnc:W_2bzt7qp7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Gemini (multiple versions), Anthropic’s Claude

View statement
Looking for related papers…
Contents