betaFeedback

WorksComputer Science

Structure-Actuation Co-Design of All-Electric Heavy-Duty Robotic Manipulators

Abstract

This paper presents a mixed discrete-continuous multiobjective bilevel framework for the structure-actuation co-design of all-electric heavy-duty robotic manipulators. The upper level optimizes manipulator geometry and actuation topology, while joint-specific lower-level searches construct feasible archives of discrete motor, gearbox, and screw configurations. A system-level coordinator selects actuator packages according to hardware mass, co-design cost, and loss-aware supply energy, and the resulting actuator masses are fed back to the robot model until a mass-consistent fixed point is obtained. Evaluation on a full-pose SE(3) benchmark with a 100-kg tool payload yields eight feasible nondominated designs, all satisfying the 12 prescribed poses. Actuator-mass coupling converges in three iterations for every reported design. A fixed-design ablation shows that one-pass sequential sizing changes hardware mass by up to 2.13%, co-design cost by 10.20%, and supply energy by 9.40% in magnitude. Payload screening further reveals increasing mass-cost-energy demand and changes in competitive mechanism topology with carried load.

Introduction

Heavy-duty robotic manipulators require high payload capacity, extended reach, and reliable operation under demanding loads, for which hydraulic actuation has traditionally been attractive [1]. Advances in electromechanical actuators (EMAs), together with the efficiency limitations of hydraulic systems, are motivating all-electric alternatives [2, 3, 4]. Electrification, however, is not simply an actuator-replacement problem. Manipulator geometry and mechanism topology determine workspace, mechanical advantage, structural loading, and actuator demand [5, 6, 7], whereas the selected motor–transmission packages determine actuator mass, reflected inertia, torque–speed capability, and efficiency [8], [9]. Since these package masses subsequently modify the robot dynamics and structural loading, mechanical design and actuator selection form a coupled structure–actuation problem rather than sequential sizing stages [10].

Integrated robot design has consequently received increasing attention. Simulation-aided and cascaded formulations have jointly optimized hardware and motion variables [11, 12, 13], while bilevel optimization has been applied to morphology–behavior coordination [14]. System-level electrical-drive optimization has likewise demonstrated the value of coordinating interacting component designs [15]. Discrete component selection [16], joint motor–motion optimization [17], linkage-energy optimization [18], and bilevel coordination of manipulator operation and EMLA efficiency [19] have further extended this direction.

This paper develops a mixed discrete–continuous multiobjective bilevel framework for structure–actuation co-design. The upper level optimizes an 11-variable mechanical design comprising link dimensions, mechanism topology, and linkage geometry, while joint-specific lower-level searches construct bounded archives of feasible EMA/EMLA packages. Their Cartesian product is coordinated at the system level according to hardware mass, co-design cost, and loss-aware supply energy. The selected actuator masses are then fed back into the rigid-body and structural models until an actuator-mass fixed point is reached. Evaluation uses a full-pose SE(3)SE(3) benchmark, a nominal co-design study, payload-sensitivity, and a fixed-design ablation of the mass-feedback loop.

The remainder of this paper is organized as follows. Section II presents the system model and benchmark, Section III describes the proposed bilevel co-design framework, Section IV discusses the results, and Section V concludes the paper.

System Model and Benchmark

Architecture and Task

The manipulator comprises a three-DoF co-designed main arm and a fixed-design three-DoF XX–YY spherical wrist. The main-arm coordinates are q=[q1,q2,q3]T\mathbf{q}=[q_1,q_2,q_3]^{\mathsf{T}}; joint 1 is rotary, whereas joints 2 and 3 may employ direct rotary actuation or three-bar or four-bar mechanisms. The wrist provides end-effector orientation but is excluded from the co-design variables.

The upper-level mechanical design is

x=[L1,L2,L3,ν2,ν3,A2,B2,α2,A3,B3,α3]T,(1)\mathbf{x}=[L_1,L_2,L_3,\nu_2,\nu_3,A_2,B_2,\alpha_2,A_3,B_3,\alpha_3]^{\mathsf{T}}, \tag*{(1)}

where νj∈{0,1,2}\nu_j\in\{0,1,2\} denotes rotary, three-bar, and four-bar mechanisms, respectively. For linearly actuated joints, AjA_j and BjB_j define the linkage geometry and αj\alpha_j provides the additional four-bar offset. Inactive variables are assigned canonical values before evaluation. The candidate joint mechanisms and their geometric parameterization are illustrated in Figure 1.

Geometric parameterization of the candidate linkage mechanisms

Figure 1. Geometric parameterization of the candidate linkage mechanisms: (a) three-bar and (b) four-bar mechanisms.

Full-pose SE(3) benchmark with 12 prescribed TCP positions and orientations and the closed dynamic-sizing trajectory

Figure 2. Full-pose SE(3)SE(3) benchmark with 12 prescribed TCP positions and orientations and the closed dynamic-sizing trajectory.

The benchmark specifies 12 full TCP poses,

Td,i=[Rd,ipd,i0T1],i=1,…,12,(2)\mathbf{T}_{d,i} = \begin{bmatrix} \mathbf{R}_{d,i} & \mathbf{p}_{d,i} \\ \mathbf{0}^{\mathsf{T}} & 1 \end{bmatrix}, \qquad i=1,\ldots,12, \tag*{(2)}

for which the complete arm–wrist chain is checked for kinematic feasibility, including joint limits, wrist singularity, solution continuity, and main-arm conditioning. For dynamic actuator sizing, the 12 TCP positions are closed into a 13-waypoint cycle and tracked using damped closed-loop inverse kinematics [20]. The wrist remains at its nominal configuration during this cycle; hence its actuator dynamics and energy are excluded. Tracking feasibility is quantified by

emax⁡=max⁡k∥pd[k]−pe[k]∥2.(3)e_{\max} = \max_k \left\lVert\mathbf{p}_d[k]-\mathbf{p}_e[k]\right\rVert_2. \tag*{(3)}

Dynamics, Linkages, and Structural Screen

For a repaired design x\mathbf{x} and actuator-mass vector ma\mathbf{m}_a, the main-arm dynamics are

τ=M(q;x,ma,md)q¨+h(q,q˙;x,ma,md),(4)\boldsymbol{\tau} = \mathbf{M}(\mathbf{q};\mathbf{x},\mathbf{m}_a,m_d)\ddot{\mathbf{q}}+\mathbf{h}(\mathbf{q},\dot{\mathbf{q}};\mathbf{x},\mathbf{m}_a,m_d), \tag*{(4)}

where md=mwrist+mtoolm_d=m_{\mathrm{wrist}}+m_{\mathrm{tool}} is the distal carried mass. The wrist contributes to hardware mass whereas the tool does not; both contribute to dynamic and structural loading.

For an EMLA-driven joint, the effective linkage angle is

qe,j={δj+qj,νj=1,δj+qj−αj,νj=2,δ2=π2, δ3=π,(5)q_{\mathrm{e},j} = \begin{cases} \delta_j+q_j, & \nu_j=1,\\ \delta_j+q_j-\alpha_j, & \nu_j=2, \end{cases} \qquad \delta_2=\frac{\pi}{2},\ \delta_3=\pi, \tag*{(5)}

and the geometry, velocity, and regularized force estimate are given below, with ϵC\epsilon_C denoting a numerical regularization parameter in metres:

Cj=Aj2+Bj2−2AjBjcos⁡qe,j,gj(qj)=AjBjsin⁡qe,jCj,va,j=gj(qj)q˙j,Fa,j=τjgj(qj)+ϵC.(6)\begin{aligned} C_j &= \sqrt{A_j^2+B_j^2-2A_jB_j\cos q_{\mathrm{e},j}},\\ g_j(q_j) &= \frac{A_jB_j\sin q_{\mathrm{e},j}}{C_j},\\ v_{a,j} &= g_j(q_j)\dot{q}_j,\qquad F_{a,j}=\frac{\tau_j}{g_j(q_j)+\epsilon_C}. \tag*{(6)} \end{aligned}

Invalid installation geometries or insufficient actuator-length margins are rejected before dynamic evaluation.

All links use a fixed hollow-square steel section with bo=0.150 mb_o=0.150\ \mathrm{m}, t=0.005 mt=0.005\ \mathrm{m}, bi=bo−2t=0.140 mb_i=b_o-2t=0.140\ \mathrm{m}, and ρ=7850 kg/m3\rho=7850\ \mathrm{kg/m^3}. Its properties are

As=bo2−bi2,Is=bo4−bi412,mL,i=ρAsLi.(7)A_s=b_o^2-b_i^2,\qquad I_s=\frac{b_o^4-b_i^4}{12},\qquad m_{L,i}=\rho A_sL_i. \tag*{(7)}

The section is fixed rather than optimized. Structural feasibility is screened using

σmax⁡=max⁡i,k(∣Fax,i[k]∣As+Mb,i[k](bo/2)Is),(8)\sigma_{\max}=\max_{i,k}\left(\frac{\lvert F_{\mathrm{ax},i}[k]\rvert}{A_s}+\frac{M_{b,i}[k](b_o/2)}{I_s}\right), \tag*{(8)}

including downstream link, actuator, wrist, and tool masses. This axial-plus-bending model is an engineering feasibility screen rather than a local-stress, fatigue, or buckling certification model.

Discrete Actuator Packages and Loss-Aware Energy

For joint jj, the lower-level decisions are yjEMA=[im,iNg]T\mathbf{y}^{\mathrm{EMA}}_j=[i_m,i_{N_g}]^{\mathsf{T}} and yjEMLA=[im,iNg,ip]T\mathbf{y}^{\mathrm{EMLA}}_j=[i_m,i_{N_g},i_p]^{\mathsf{T}}, selecting a catalogue motor, gearbox ratio NgN_g, and, for an EMLA, screw lead pp. With NgN_g denoting the motor-to-output speed ratio, ωm,j=Ngq˙j\omega_{m,j}=N_g\dot{q}_j for EMAs and ωm,j=(2πNg/p)va,j\omega_{m,j}=(2\pi N_g/p)v_{a,j} for EMLAs, where pp is expressed in m/rev. Complete variable package mass and cost are

ma,j=mm+mg+IEMLAms+maux,Ca,j=Cm+Cg+IEMLACs+Caux,(9)m_{a,j}=m_m+m_g+\mathbb{I}_{\mathrm{EMLA}}m_s+m_{\mathrm{aux}}, \qquad C_{a,j}=C_m+C_g+\mathbb{I}_{\mathrm{EMLA}}C_s+C_{\mathrm{aux}}, \tag*{(9)}

including auxiliary hardware. Motor torque, speed, power, and current limits are enforced, while EMLA screw feasibility also includes axial strength, buckling, and critical speed [21].

Supply energy is estimated using a catalogue-calibrated load-dependent loss surrogate,

Ploss,N=PN(ηN−1−1),P_{\mathrm{loss,N}}=P_N(\eta_N^{-1}-1),
Ploss(t)=Ploss,N(τm(t)τN)2,(10)P_{\mathrm{loss}}(t)=P_{\mathrm{loss,N}}\left(\frac{\tau_m(t)}{\tau_N}\right)^2, \tag*{(10)}
Esup=∫max⁡[Pm(t)+Ploss(t),0] dt,E_{\mathrm{sup}}=\int\max\left[P_m(t)+P_{\mathrm{loss}}(t),0\right]\,\mathrm{d}t,

@page 2/6 starts_mid=0 ends_mid=1

Geometric parameterization of the candidate linkage mechanisms

Figure 1 (PDF p. 2). Geometric parameterization of the candidate linkage mechanisms: (a) three-bar and (b) four-bar mechanisms.

Full-pose SE(3) benchmark with 12 prescribed TCP positions and orientations and the closed dynamic-sizing trajectory

Figure 2 (PDF p. 2). Full-pose SE(3)SE(3) benchmark with 12 prescribed TCP positions and orientations and the closed dynamic-sizing trajectory.

The benchmark specifies 12 full TCP poses,

Td,i=[Rd,ipd,i0T1],i=1,…,12,(11)\mathbf{T}_{d,i} = \begin{bmatrix} \mathbf{R}_{d,i} & \mathbf{p}_{d,i} \\ \mathbf{0}^{\mathsf{T}} & 1 \end{bmatrix}, \qquad i=1,\ldots,12, \tag*{(11)}

for which the complete arm–wrist chain is checked for kinematic feasibility, including joint limits, wrist singularity, solution continuity, and main-arm conditioning. For dynamic actuator sizing, the 12 TCP positions are closed into a 13-waypoint cycle and tracked using damped closed-loop inverse kinematics [20]. The wrist remains at its nominal configuration during this cycle; hence its actuator dynamics and energy are excluded. Tracking feasibility is quantified by

emax⁡=max⁡k∥pd[k]−pe[k]∥2.(12)e_{\max} = \max_k \left\lVert\mathbf{p}_d[k]-\mathbf{p}_e[k]\right\rVert_2. \tag*{(12)}

Dynamics, Linkages, and Structural Screen

For a repaired design x\mathbf{x} and actuator-mass vector ma\mathbf{m}_a, the main-arm dynamics are

τ=M(q;x,ma,md)q¨+h(q,q˙;x,ma,md),(13)\boldsymbol{\tau} = \mathbf{M}(\mathbf{q};\mathbf{x},\mathbf{m}_a,m_d)\ddot{\mathbf{q}}+\mathbf{h}(\mathbf{q},\dot{\mathbf{q}};\mathbf{x},\mathbf{m}_a,m_d), \tag*{(13)}

where md=mwrist+mtoolm_d=m_{\mathrm{wrist}}+m_{\mathrm{tool}} is the distal carried mass. The wrist contributes to hardware mass whereas the tool does not; both contribute to dynamic and structural loading.

For an EMLA-driven joint, the effective linkage angle is

qe,j={δj+qj,νj=1,δj+qj−αj,νj=2,δ2=π2, δ3=π,(14)q_{\mathrm{e},j} = \begin{cases} \delta_j+q_j, & \nu_j=1,\\ \delta_j+q_j-\alpha_j, & \nu_j=2, \end{cases} \qquad \delta_2=\frac{\pi}{2},\ \delta_3=\pi, \tag*{(14)}

and the geometry, velocity, and regularized force estimate are given below, with ϵC\epsilon_C denoting a numerical regularization parameter in metres:

Cj=Aj2+Bj2−2AjBjcos⁡qe,j,gj(qj)=AjBjsin⁡qe,jCj,va,j=gj(qj)q˙j,Fa,j=τjgj(qj)+ϵC.(15)\begin{aligned} C_j &= \sqrt{A_j^2+B_j^2-2A_jB_j\cos q_{\mathrm{e},j}},\\ g_j(q_j) &= \frac{A_jB_j\sin q_{\mathrm{e},j}}{C_j},\\ v_{a,j} &= g_j(q_j)\dot{q}_j,\qquad F_{a,j}=\frac{\tau_j}{g_j(q_j)+\epsilon_C}. \tag*{(15)} \end{aligned}

Invalid installation geometries or insufficient actuator-length margins are rejected before dynamic evaluation.

All links use a fixed hollow-square steel section with bo=0.150 mb_o=0.150\ \mathrm{m}, t=0.005 mt=0.005\ \mathrm{m}, bi=bo−2t=0.140 mb_i=b_o-2t=0.140\ \mathrm{m}, and ρ=7850 kg/m3\rho=7850\ \mathrm{kg/m^3}. Its properties are

As=bo2−bi2,Is=bo4−bi412,mL,i=ρAsLi.(16)A_s=b_o^2-b_i^2,\qquad I_s=\frac{b_o^4-b_i^4}{12},\qquad m_{L,i}=\rho A_sL_i. \tag*{(16)}

The section is fixed rather than optimized. Structural feasibility is screened using

σmax⁡=max⁡i,k(∣Fax,i[k]∣As+Mb,i[k](bo/2)Is),(17)\sigma_{\max}=\max_{i,k}\left(\frac{\lvert F_{\mathrm{ax},i}[k]\rvert}{A_s}+\frac{M_{b,i}[k](b_o/2)}{I_s}\right), \tag*{(17)}

including downstream link, actuator, wrist, and tool masses. This axial-plus-bending model is an engineering feasibility screen rather than a local-stress, fatigue, or buckling certification model.

Discrete Actuator Packages and Loss-Aware Energy

For joint jj, the lower-level decisions are yjEMA=[im,iNg]T\mathbf{y}^{\mathrm{EMA}}_j=[i_m,i_{N_g}]^{\mathsf{T}} and yjEMLA=[im,iNg,ip]T\mathbf{y}^{\mathrm{EMLA}}_j=[i_m,i_{N_g},i_p]^{\mathsf{T}}, selecting a catalogue motor, gearbox ratio NgN_g, and, for an EMLA, screw lead pp. With NgN_g denoting the motor-to-output speed ratio, ωm,j=Ngq˙j\omega_{m,j}=N_g\dot{q}_j for EMAs and ωm,j=(2πNg/p)va,j\omega_{m,j}=(2\pi N_g/p)v_{a,j} for EMLAs, where pp is expressed in m/rev. Complete variable package mass and cost are

ma,j=mm+mg+IEMLAms+maux,Ca,j=Cm+Cg+IEMLACs+Caux,(18)m_{a,j}=m_m+m_g+\mathbb{I}_{\mathrm{EMLA}}m_s+m_{\mathrm{aux}}, \qquad C_{a,j}=C_m+C_g+\mathbb{I}_{\mathrm{EMLA}}C_s+C_{\mathrm{aux}}, \tag*{(18)}

including auxiliary hardware. Motor torque, speed, power, and current limits are enforced, while EMLA screw feasibility also includes axial strength, buckling, and critical speed [21].

Supply energy is estimated using a catalogue-calibrated load-dependent loss surrogate,

Ploss,N=PN(ηN−1−1),P_{\mathrm{loss,N}}=P_N(\eta_N^{-1}-1),
Ploss(t)=Ploss,N(τm(t)τN)2,(19)P_{\mathrm{loss}}(t)=P_{\mathrm{loss,N}}\left(\frac{\tau_m(t)}{\tau_N}\right)^2, \tag*{(19)}
Esup=∫max⁡[Pm(t)+Ploss(t),0] dt,E_{\mathrm{sup}}=\int\max\left[P_m(t)+P_{\mathrm{loss}}(t),0\right]\,\mathrm{d}t,

where Pm=τmωmP_m=\tau_m\omega_m is signed mechanical motor power. Negative net power is not credited as regeneration. The model is intended for comparative actuator selection; speed-dependent iron, windage, inverter, and thermal losses are not explicitly resolved. Wrist-actuator energy is likewise excluded because the timed dynamic model considers only the co-designed main arm.

Mass-Consistent Bilevel Co-Design

The mechanical design is the leader decision, while actuator selection provides its mass-consistent follower response:

minimize⁡x∈XFu(x,yb⋆(x))subject toGu(x,yb⋆(x))⪯0,(20)\begin{aligned} \underset{x\in\mathcal{X}}{\operatorname{minimize}}\quad&F_u(x,y_b^\star(x))\\ \text{subject to}\quad&G_u(x,y_b^\star(x))\preceq0, \tag*{(20)} \end{aligned}

where yb⋆y_b^\star is the converged three-joint actuator package obtained through joint-level search, system-level coordination, and actuator-mass feedback. Hence, each leader candidate is evaluated under the mass distribution induced by its selected actuator packages.

For given (x,ma)(x,m_a), the robot simulation provides the joint histories used for actuator selection. The family-dependent lower objectives are

Fℓ,j={[pNkW,Ng,ρJ,j]T,EMA,[pNkW,Ng,pmm,ρm,j]T,EMLA,(21)F_{\ell,j}= \begin{cases} \left[p_N^{\mathrm{kW}},N_g,\rho_{J,j}\right]^{\mathrm{T}}, & \mathrm{EMA},\\ \left[p_N^{\mathrm{kW}},N_g,p^{\mathrm{mm}},\rho_{m,j}\right]^{\mathrm{T}}, & \mathrm{EMLA}, \end{cases} \tag*{(21)}

where

{ρJ,j=JmNg2Jeq,jmin⁡,+,ρm,j=Jm(2πNg/p)2meq,jmin⁡,+.(22)\begin{cases} \rho_{J,j}=\dfrac{J_mN_g^2}{J_{\mathrm{eq},j}^{\min,+}},\\ \rho_{m,j}=\dfrac{J_m(2\pi N_g/p)^2}{m_{\mathrm{eq},j}^{\min,+}}. \tag*{(22)} \end{cases}

Here JmJ_m is motor inertia, while Jeq,jmin⁡,+J_{\mathrm{eq},j}^{\min,+} and meq,jmin⁡,+m_{\mathrm{eq},j}^{\min,+} are the minimum positive equivalent rotational inertia and translational mass over the task. The ratios discourage excessive reflected motor inertia.

Subject to the actuator and topology-dependent constraints, the feasible discrete follower set is

Yℓ,jfeas(x,ma)={yj∈Yj(x):Gℓ,j(x,ma,yj)⪯0}.(23)\mathcal{Y}_{\ell,j}^{\mathrm{feas}}(x,m_a)=\left\{y_j\in\mathcal{Y}_j(x):G_{\ell,j}(x,m_a,y_j)\preceq0\right\}. \tag*{(23)}

Rather than returning a single joint solution, each search retains a bounded archive Aj(x,ma)⊂Yℓ,jfeas\mathcal{A}_j(x,m_a)\subset\mathcal{Y}_{\ell,j}^{\mathrm{feas}}. Nondominated solutions are prioritized while additional feasible alternatives are retained for system-level coordination.

System-Level Package Coordination and Mass Feedback

The joint archives are coordinated jointly rather than selecting each actuator independently. Omitting their dependence on (x,ma)(x,m_a) for brevity, define

{A=A1×A2×A3,z(y)=[Ma(y),Ca(y),Esup(y)]T,y∈A.(24)\begin{cases} \mathcal{A}=\mathcal{A}_1\times\mathcal{A}_2\times\mathcal{A}_3,\\ z(y)=\left[M_a(y),C_a(y),E_{\mathrm{sup}}(y)\right]^{\mathrm{T}},\qquad y\in\mathcal{A}. \tag*{(24)} \end{cases}

where each yy is a realizable three-joint package and MaM_a, CaC_a, and EsupE_{\mathrm{sup}} are its total actuator mass, cost, and supply energy.

The three metrics are min–max normalized over A\mathcal{A}, and an equal-weight ideal-point compromise is selected:

{z^r(y)=zr(y)−zrmin⁡zrmax⁡−zrmin⁡+ϵz,r=1,2,3,yb(k)=arg⁡min⁡y∈A∥z^(y)∥2.(25)\begin{cases} \hat{z}_r(y)=\dfrac{z_r(y)-z_r^{\min}}{z_r^{\max}-z_r^{\min}+\epsilon_z},\qquad r=1,2,3,\\ y_b^{(k)}=\underset{y\in\mathcal{A}}{\arg\min}\left\|\hat{z}(y)\right\|_2. \tag*{(25)} \end{cases}

Thus, mass, cost, and energy enter the coordination with equal explicit weighting after normalization.

For fixed xx, robot simulation, joint-level search, and package coordination define the actuator-mass response Tx\mathcal{T}_x. The coupling iteration is

{ma(k+1)=Tx(ma(k))=ma(yb(k)),em(k)=∥ma(k+1)−ma(k)∥2∥ma(k)∥2+ϵ.(26)\begin{cases} m_a^{(k+1)}=\mathcal{T}_x(m_a^{(k)})=m_a(y_b^{(k)}),\\ e_m^{(k)}=\dfrac{\left\|m_a^{(k+1)}-m_a^{(k)}\right\|_2}{\left\|m_a^{(k)}\right\|_2+\epsilon}. \tag*{(26)} \end{cases}

If em(k)>ϵme_m^{(k)}> \epsilon_m, the updated masses are returned to the robot model and the actuator demands and package selection are recomputed. Iteration terminates at em(k)≤ϵme_m^{(k)}\leq\epsilon_m, yielding the converged package yb⋆y_b^\star and mass vector ma⋆≈Tx(ma⋆)m_a^\star\approx\mathcal{T}_x(m_a^\star). The upper-level objectives and constraints are then evaluated using these mass-consistent installed actuator masses.

Upper-Level Objectives and Feasibility

After actuator-mass convergence, the upper objectives are

Fu(x,yb⋆)=[MhwCcodesignEsup]=[Mlink+Ma+MwristCstructure+CaEsup].(27)F_u(x,y_b^\star)= \begin{bmatrix} M_{\mathrm{hw}}\\ C_{\mathrm{codesign}}\\ E_{\mathrm{sup}} \end{bmatrix} = \begin{bmatrix} M_{\mathrm{link}}+M_a+M_{\mathrm{wrist}}\\ C_{\mathrm{structure}}+C_a\\ E_{\mathrm{sup}} \end{bmatrix}. \tag*{(27)}

The tool payload is excluded from MhwM_{\mathrm{hw}} but retained in the dynamic and structural loading. Likewise, the fixed wrist cost is omitted from CcodesignC_{\mathrm{codesign}} because it is common to all designs.

A candidate is feasible only if the converged system satisfies the prescribed SE(3)SE(3) task, kinematic and geometric requirements, dynamic tracking and structural-stress limits, actuator ratings, topology-dependent EMLA constraints, and actuator-mass consistency. Constraint handling uses feasible-first constrained Pareto dominance: feasible designs are ranked by their objective vectors, whereas infeasible designs are ordered by aggregate constraint violation. The upper search therefore returns a finite nondominated approximation of mass-consistent structure–actuation designs, rather than a scalar optimum. Both upper- and lower-level multiobjective searches employ NSGA-II [22].

Results and Discussion

Numerical Setup

The nominal study considers a 100 kg tool payload and a fixed 50 kg spherical wrist, giving a distal carried mass of 150 kg in the dynamic and structural models. Table I summarizes the mechanical search space and numerical settings. The actuator catalogue contains 68 motor records, while gearbox ratios and EMLA screw leads are discrete. Full-pose feasibility requires all 12 prescribed SE(3)SE(3) targets, a minimum joint-limit and wrist-singularity margin of 5∘5^\circ, and a

ItemValue
L1L_1; L2L_2, L3L_3[0.8, 2.0] m; [2.5, 5.0] m
ν2\nu_2, ν3\nu_3rotary / 3-bar / 4-bar
A2A_2, A3A_3; B2B_2, B3B_3[0.60, 0.90] m; [0.30, 0.90] m
α2\alpha_2; α3\alpha_3[0.01, 0.50] rad; [0.01, 0.20] rad
Motor / NN / pp68 motors; N,p∈{1,…,30}N,p \in\{1,\ldots,30\}, pp in mm/rev
Tool / fixed wrist100 kg / 50 kg
Upper NSGA-IIpopulation 20; generation cap 15; seed 101
Lower NSGA-IIpopulation 18; generation cap 12; seed 1001
Joint archiveat most 12 retained alternatives per joint
Mass couplingma(0)=[15,15,15]Tm_a^{(0)} = [15,15,15]^{\mathsf{T}} kg; Kmax⁡=6K_{\max}=6; ϵm=5×10−3\epsilon_m=5 \times10^{-3}
Simulationfull Δt\Delta t scale 1.0; emax⁡≤20e_{\max} \leq20 mm; σmax⁡≤250\sigma_{\max} \leq250 MPa

Table I. Principal Design and Optimization Settings

Algorithm 1 Multiobjective Bilevel Structure–Actuation Co-Design

Require: Mechanical bounds, actuator catalogue, task data, ma(0)m_a^{(0)}, Kmax⁡K_{\max}, ϵm\epsilon_m

  1. Initialize a repaired upper-level population

  2. for each upper NSGA-II generation do

    1. for each unevaluated mechanical design x\boldsymbol{x} do

      1. Repair x\boldsymbol{x} and canonicalize inactive variables

      2. ma←ma(0)m_a \leftarrow m_a^{(0)}

      3. for k=0,…,Kmax⁡−1k = 0,\ldots,K_{\max}-1 do

        1. Simulate robot using (x,ma)(\boldsymbol{x},m_a) and extract joint demand histories

        2. for j=1,2,3j = 1,2,3 do

          1. Run topology-dependent constrained lower NSGA-II and construct AjA_j

        3. end for

        4. Form A1×A2×A3A_1 \times A_2 \times A_3 and select yb(k)y_b^{(k)} using the normalized mass–cost–energy ideal point

        5. ma(k+1)←ma(yb(k))m_a^{(k+1)} \leftarrow m_a(y_b^{(k)})

        6. if em(k)≤ϵme_m^{(k)} \leq\epsilon_m then

          1. break

        7. end if

        8. ma←ma(k+1)m_a \leftarrow m_a^{(k+1)}

      4. end for

      5. Perform final package-mass-consistent evaluation

      6. Evaluate FuF_u and GuG_u

    2. end for

    3. Apply feasible-first nondominated sorting, crowding, selection, and variation

  3. end for

  4. return Obtained feasible rank-zero designs

main-arm condition number not exceeding 100. The dynamic and structural limits are emax⁡≤20e_{\max} \leq20 mm and σmax⁡≤250\sigma_{\max} \leq250 MPa, respectively.

For reporting, the co-design cost is normalized as

Cidx=CcodesignCref,(28)C_{\mathrm{idx}} = \frac{C_{\mathrm{codesign}}}{C_{\mathrm{ref}}}, \tag*{(28)}

where CrefC_{\mathrm{ref}} is the lowest co-design cost observed across the nominal and payload-screening datasets. The same reference is used for all reported cost indices. The nominal campaign uses the full simulation resolution (Δt\Delta t scale 1.0) and fixed upper- and lower-level random seeds for reproducibility.

Nominal 100-kg Tool-Payload Co-Design

The fixed-budget nominal search evaluated 320 unique upper-level designs and returned eight feasible nondominated solutions satisfying all kinematic, dynamic, structural, actuator, and mass-consistency requirements. Figure 3 shows their three-objective trade-offs.

Figure 3. Nominal 100-kg nondominated set: (a) hardware mass versus normalized co-design cost, (b) hardware mass versus supply energy, and (c) normalized cost versus supply energy. DMD_M, DCD_C, and DED_E denote the minimum-mass, minimum-cost, and minimum-energy designs.

The set spans 387.62–419.74 kg in hardware mass, 1.187–1.310 in normalized co-design cost, and 30.53–36.90 kJ in supply energy. The extrema are

MhwM_{\mathrm{hw}} [kg]CidxC_{\mathrm{idx}}EsupE_{\mathrm{sup}} [kJ]
DMD_M387.621.20732.95
DCD_C419.741.18736.90
DED_E391.131.31030.53

showing that improvement of one objective generally requires sacrifice in another.

Table II exposes the structure–actuation coupling behind these objective extremes. Although DMD_M and DED_E share the same 4-bar/3-bar topology and principal link dimensions, their linkage geometry and drive packages differ markedly: DMD_M uses an 11-kW, Ng=15N_g=15 motor at J1J_1 and a 4-kW motor at J2J_2, whereas DED_E uses a 3-kW, Ng=26N_g=26 motor at J1J_1 and an 11-kW motor at J2J_2. Their J2J_2 selections remain nearly unchanged. In contrast, DCD_C adopts a 4-bar/4-bar topology and 2.2-kW motors at both linearly actuated joints. Thus, similar arm dimensions can yield substantially different electromechanical realizations, confirming that the trade-off is governed jointly by mechanical architecture and discrete actuator selection.

Payload Sensitivity

Payload sensitivity was screened at tool loads of 20, 60, 100, and 140 kg using a common reduced fixed search budget and the fixed 50-kg wrist. These results therefore indicate payload-dependent trends rather than converged payload-specific optima.

As the tool payload increases from 20 to 140 kg, the minimum observed hardware mass rises from 348.87 to 442.16 kg, the cost index from 1.000 to 1.394, and supply energy from 22.86 to 40.33 kJ. The competitive mechanism topology also changes with load: the 20-kg set contains 3-bar/4-bar and 4-bar/4-bar designs; the 60- and 100-kg sets contain 3-bar/4-bar, 4-bar/3-bar, and 4-bar/4-bar designs; and the 140-kg set retains only 3-bar/4-bar and 4-bar/3-bar designs. Moreover, the minimum-mass, minimum-cost, and minimum-energy solutions need not share the same topology, showing that payload influences both actuator demand and the preferred mechanical architecture.

Effect of Actuator-Mass Feedback

Actuator-mass coupling was isolated through a fixed-design ablation of DMD_M, DED_E, and DCD_C. For each design, actuators were sized once from the common initial estimate ma(0)=[15,15,15]Tm_a^{(0)} = [15,15,15]^{\mathsf{T}} kg; the selected package was then frozen and re-evaluated using its installed masses without further actuator AjA_j, BjB_j are in m and αj\alpha_j in rad. M01–M07 denote anonymized catalogue motors; PNP_N is in kW, NgN_g is the gearbox ratio, and pp is the EMLA screw lead in mm/rev.

DesignJ2/J3J_2/J_3[L1,L2,L3][L_1,L_2,L_3] [m][A2,B2,α2][A_2,B_2,\alpha_2][A3,B3,α3][A_3,B_3,\alpha_3]J1J_1: M/PN/NgM/P_N/N_gJ2J_2: M/PN/Ng/pM/P_N/N_g/pJ3J_3: M/PN/Ng/pM/P_N/N_g/p
DMD_M4-bar/3-bar[1.145, 4.961, 2.543][0.772, 0.370, 0.329][0.817, 0.678, 0.100]M01/11/15M02/4/6/7M03/2.2/3/5
DED_E4-bar/3-bar[1.145, 4.961, 2.543][0.683, 0.828, 0.060][0.817, 0.699, 0.100]M04/3/26M01/11/6/7M03/2.2/3/5
DCD_C4-bar/4-bar[1.213, 4.961, 2.682][0.683, 0.828, 0.100][0.654, 0.397, 0.100]M05/5.5/15M06/2.2/9/13M07/2.2/6/9

Table 2. Representative nominal co-designs.

Payload screening

Figure 4. Payload screening: minimum observed (a) hardware mass, (b) normalized co-design cost, and (c) supply energy within each obtained nondominated set. All cases use the same reduced fixed search budget.

Designem(0)e_m^{(0)}ΔM\Delta M [%]ΔC\Delta C [%]ΔEsup\Delta E_{\mathrm{sup}} [%]ϵE\epsilon_E [%]
DMD_M2.655+0.47-0.71+9.40+1.89
DED_E2.836+2.13-3.81+8.09+1.75
DCD_C3.367+0.43+10.20-8.80+3.67

Table 3. Fixed-design actuator-mass-feedback ablation.

reselection. This one-pass baseline was compared with the fully coupled solution using

Δz=100z~1pass−z~coupledz~coupled,ϵE=100Esup,installed−Esup,predEsup,pred.(29)\Delta z = 100\frac{\tilde{z}_{\mathrm{1pass}}-\tilde{z}_{\mathrm{coupled}}}{\tilde{z}_{\mathrm{coupled}}}, \qquad \epsilon_E = 100\frac{E_{\mathrm{sup,installed}}-E_{\mathrm{sup,pred}}}{E_{\mathrm{sup,pred}}}. \tag*{(29)}

The first-update discrepancies, em(0)=2.655e_m^{(0)}=2.655–3.3673.367, quantify differences between the initial estimates and selected package masses. All frozen one-pass packages satisfy the physical constraints after re-evaluation with their installed masses. Actuator reselection changes the realized trade-off: hardware mass changes by 0.43–2.13%, co-design cost by −3.81 to +10.20%, and supply energy by −8.80 to +9.40%. The discrete follower response changes as well; for example, DMD_M changes its J2J_2 screw lead from 10 to 7 mm/rev after coupling, while DED_E changes both its J1J_1 package and J2J_2 screw selection. Thus, actuator-mass feedback makes actuator selection consistent with the demand histories induced by the selected packages, although it does not systematically improve every individual objective.

For DMD_M, the mass iteration evolves from [15,15,15]T[15,15,15]^{\mathrm{T}} to [77.36,43.72,21.49]T[77.36,43.72,21.49]^{\mathrm{T}} and then [77.86,41.40,21.49]T[77.86,41.40,21.49]^{\mathrm{T}} kg, which is reproduced on the following iteration. The remaining reported nominal designs likewise satisfy the mass-convergence criterion in three iterations.

Constraint and Physical Validation

All eight nominal nondominated designs satisfy the prescribed physical constraints and all 12 full-pose targets. Across the set, the worst-case tracking error and structural stress are 9.73 mm and 128.35 MPa, below their respective limits of 20 mm and 250 MPa. The minimum joint-limit and EMLA stroke margins are 48.05° and 34.41 mm, while the maximum main-arm condition number is 6.11. Maximum torque, speed, and power utilizations are 0.995, 0.983, and 0.946, respectively. Actuator-mass coupling converges for every reported design in three iterations with zero final residual and no detected cycle.

Because the evolutionary search uses a prescribed finite budget, the reported designs constitute a nondominated approximation rather than a globally converged Pareto front.

Conclusion

A multiobjective bilevel structure–actuation co-design framework was developed for all-electric heavy-duty robotic manipulators. Mixed mechanism topology and continuous geometry are optimized jointly with discrete actuator packages, while system-level package coordination and actuator-mass feedback ensure consistency between actuator sizing and robot dynamics.

For the nominal 100-kg tool-payload case, the fixed-budget search produced eight feasible nondominated designs satisfying all 12 prescribed SE(3)SE(3) poses, with maximum structural stress of 128.35 MPa. Actuator-mass coupling converged for every reported design in three iterations. The fixed-design ablation showed that one-pass sequential sizing changes hardware mass by up to 2.13%, co-design cost by 10.20%, and supply energy by 9.40% in magnitude, demonstrating that actuator-mass feedback can alter both the selected discrete package and the resulting system trade-off. Payload screening further showed increasing mass–cost–energy demand and changes in competitive mechanism topology with carried load.

The reported solutions constitute a fixed-budget nondominated approximation rather than a globally converged Pareto front. Current limitations include the engineering structural and motor-loss models, the fixed-design wrist, and omission of wrist-actuator energy. Future work will address higher-fidelity component and thermal models, robust co-design under uncertainty, and experimental validation.

References

  1. [1]G. R. Petrović and J. Mattila, “Mathematical modelling and virtual decomposition control of heavy-duty parallel–serial hydraulic manipulators,” Mechanism and Machine Theory, vol. 170, p. 104680, 2022.DOI
  2. [2]Z. Li, C. Wang, L. Quan, Y. Hao, L. Ge, and L. Xia, “Study on energy efficiency characteristics of the heavy-duty manipulator driven by electro-hydraulic hybrid active-passive system,” Automation in Construction, vol. 125, p. 103646, 2021.DOI
  3. [3]M. Bahari, P. Mustalahti, Á. P. Anaya, and J. Mattila, “Electrification of heavy-duty robotic manipulators with pmsm-driven electromechanical linear actuators,” IEEE Access, vol. 14, pp. 21 805–21 832, 2026.DOI
  4. [4]M. Bahari, A. H. Barjini, P. Mustalahti, and J. Mattila, “All-electric heavy-duty robotic manipulator: Actuator configuration optimization and sensorless control,” arXiv preprint arXiv:2509.15778, 2025.arxiv.org/abs/2509.15778
  5. [5]R. Wu, W. Yang, and S. Bai, “Exact solutions of the mixed motion and path synthesis problem for four-bar linkages,” Journal of Mechanisms and Robotics, vol. 17, no. 7, p. 071005, 2025.DOI
  6. [6]S. N. Nabavi, M. Sharatiee, J. Enferadi, and A. Akbarzadeh, “Parametric design and multi-objective optimization of a general 6-pus parallel manipulator,” Mechanism and Machine Theory, vol. 152, p. 103913, 2020.DOI
  7. [7]X. Wu, Q. Liu, J. Ding, C. Wang, H. Yu, and S. Bai, “Transmission angle of planar four-bar linkages applicable for different input-output links subject to external loads,” Mechanism and Machine Theory, vol. 203, p. 105829, 2024.DOI
  8. [8]I. H. Park and D. S. Hong, “Selection of motors and gear reducers for four-axis palletizing robots considering the motion characteristics and dynamic coupling effect,” International Journal of Precision Engineering and Manufacturing, vol. 24, no. 12, pp. 2269–2278, 2023.DOI
  9. [9]M. Bahari, A. Paz, A. S. Habib, and J. Mattila, “Performance evaluation of an electromechanical linear actuator with optimal trajectories,” in 2023 IEEE 97th Vehicular Technology Conference (VTC2023-Spring), 2023, pp. 1–7.DOI
  10. [10]Z. Lu and Y. Wang, “A differentiable dynamic modeling approach to integrated motion planning and actuator physical design for mobile manipulators,” Journal of Field Robotics, vol. 42, no. 1, pp. 37–64, 2025.arxiv.org/abs/2405.00882
  11. [11]G. Fadini, T. Flaylos, A. Del Prete, and P. Souderes, “Simulation aided co-design for robust robot optimization,” IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 11 306–11 313, 2022.DOI
  12. [12]H. Peralta-Caprachin, R. Angeles-Orahulio, and E. Paiva-Peredo, “Design and position control of a robot with 5 degrees of freedom,” International Journal of Mechanical Engineering and Robotics Research, vol. 13, no. 2, pp. 241–248, 2024.
  13. [13]A. H. Barjini, M. Bahari, M. Hejrati, and J. Mattila, “Surrogate-enhanced modeling and adaptive modular control of all-electric heavy-duty robotic manipulators,” arXiv preprint arXiv:2508.06313, 2025.arxiv.org/abs/2508.06313
  14. [14]Y. Kim, Z. Pan, and K. Hauser, “Mo-bbo: Multi-objective bilevel bayesian optimization for robot and behavior co-design,” in 2021 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2021, pp. 9877–9884.DOI
  15. [15]G. Lei, T. Wang, Y. Guo, J. Zhu, and S. Wang, “System-level design optimization methods for electrical drive systems: Deterministic approach,” IEEE Transactions on Industrial Electronics, vol. 61, no. 12, pp. 6591–6602, 2014.DOI
  16. [16]E. M. Hoffmann, D. Costanzi, G. Fadini, N. Miguel, A. Del Prete, and L. Marchionni, “Addressing reachability and discrete component selection in robotic manipulator design through kineto-static bi-level optimization,” IEEE Robotics and Automation Letters, vol. 10, no. 3, pp. 2263–2270, 2025.
  17. [17]A. Stein, Y. Wang, Y. Sakamoto, B. Wang, and H. Fang, “Application-oriented co-design of motors and motions for a 6dof robot manipulator,” in 2025 IEEE 21st International Conference on Automation Science and Engineering (CASE). IEEE, 2025, pp. 3318–3324.arxiv.org/abs/2310.03132
  18. [18]A. Paz, M. Bahari, and J. Mattila, “Energy-cautious designation of kinematic parameters for a sustainable parallel-serial heavy-duty manipulator driven by electromechanical linear actuator,” in 2024 IEEE 100th Vehicular Technology Conference (VTC2024-Fall). IEEE, 2024, pp. 1–7.arxiv.org/abs/2410.08600
  19. [19]M. Bahari, A. Paz Anaya, M. Heydari Shahna, P. Mustalahti, and J. Mattila, “System-level efficient performance of emla-driven heavy-duty manipulators via bilevel optimization framework with a leader-follower scenario,” IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 16 373–16 396, 2025.DOI
  20. [20]S. Yu and G. Tan, “Inverse kinematics of a 7-degree-of-freedom robotic arm based on deep reinforcement learning and damped least squares,” IEEE Access, vol. 13, pp. 4857–4868, 2024.
  21. [21]R. C. Jang, M. I. Kim, and K. S. Ri, “Selection of key components in ball screw-driven servomechanisms with toothed belt drive transmission for machine tools through combinational optimization,” The International Journal of Advanced Manufacturing Technology, vol. 138, no. 9, pp. 4063–4080, 2025.DOI
  22. [22]K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: Nsga-ii,” IEEE transactions on evolutionary computation, vol. 6, no. 2, pp. 182–197, 2002.DOI

Paper details

vlnc:W_apg5buv3
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents