Flexible Cable-Driven Robotic Arms (FCRAs)

• Flexible Cable-Driven Robotic Arms (FCRAs) are robotic manipulators actuated by cables instead of rigid gear trains, offering low inertia, remote actuation, and enhanced compliance.
• They encompass diverse architectures—from modular soft arms to rigid-link and hybrid systems—with innovative cable routing that decouples actuation for precise control.
• Recent studies demonstrate improved workspace scalability and dynamic performance, while addressing challenges like cable elasticity, hysteresis, and motion coupling.

Searching arXiv for recent and foundational papers on flexible cable-driven robotic arms and closely related cable-driven robot architectures. arxiv_search(query="Flexible Cable-Driven Robotic Arms cable-driven soft robotic arm modular reconfigurable end-effector cable robot redundant manipulator", max_results=10) arxiv_search(query="(Liang et al., 21 Jul 2025) OR (Islam et al., 2 Mar 2026) OR (Luo et al., 18 Feb 2025) OR (Qi et al., 2024) OR (Pornthisan et al., 3 Jun 2026) OR (Zhang, 26 Oct 2025) OR (Bury et al., 2019) OR (Qi et al., 2019) OR (Mahapatra et al., 2020)", max_results=10) Flexible cable-driven robotic arms (FCRAs) are robotic manipulators whose joints are actuated through cables rather than rigid gear trains or direct-drive motors. Across recent work, the term encompasses soft tendon-actuated continuum arms, fully decoupled rigid-link cable-driven arms, cable-driven redundant manipulators, and hybrid systems in which a serial arm is mounted on a cable-driven parallel robot. The recurring motivation is consistent: cable transmission enables remote actuation, low inertia at the moving structure, backlash-free transmission, compliance, and operation in constrained or harsh environments, but it also introduces resilience, hysteresis, friction, motion coupling, routing complexity, cable slack, non-unique tension allocation, and collision-checking difficulties (Liang et al., 21 Jul 2025, Luo et al., 18 Feb 2025, Qi et al., 2019).

1. Scope and canonical system classes

The literature represented here suggests that FCRAs should not be treated as a single morphology. One established class is the modular soft arm built from repeatable compliant sections. In "A Novel Modular Cable-Driven Soft Robotic Arm with Multi-Segment Reconfigurability" (Islam et al., 2 Mar 2026), each segment is soft and compliant, mechanically stackable, actuated by three tendons, and controlled by its own motor set. In "Design and Nonlinear Modeling of a Modular Cable Driven Soft Robotic Arm" (Qi et al., 2024), each section is a hybrid soft–rigid module made of a soft tubing backbone, a soft silicone arm body, and two rigid endcaps that connect adjacent sections and decouple the actuation cables of different sections.

A second class is the rigid-link, remotely actuated cable-driven arm with explicit transmission decoupling. D3-Arm is a 6-DOF cable-driven robotic arm in which all motors and electrical components are placed at the base, the moving part is 776 mm long, and the moving mass is 1.6 kg; its architecture combines six pairs of drive cables, 1-DOF decoupling cable aligner mechanisms, 1-DOF decoupling rolling pair mechanisms, and a cable-pretension module (Luo et al., 18 Feb 2025).

A third class is the cable-driven redundant or continuum manipulator whose geometry itself is part of the design problem. The quaternion-joint cable-driven redundant manipulator of (Pornthisan et al., 3 Jun 2026) uses a 4-segment, 8-joint configuration, whereas the 3D-printed continuum robot of (Mahapatra et al., 2020) studies six different cable routings, including straight, helical, and arbitrary general routings.

A fourth class extends the notion of an arm to hybrid cable-driven systems. In (Qi et al., 2019), the generalized flexible Hybrid Cable-Driven Robot (HCDR) is defined as two or more heterogeneous mechatronic components, where at least one component is a cable-driven parallel robot; the concrete architecture is a cable-driven mobile platform carrying a serial robot arm. In (Bury et al., 2019), the same hybridization appears in planning form: a cable-driven parallel robot with an embarked robotic arm.

A common misconception is that FCRAs are synonymous with monolithic soft manipulators. The cited systems include monolithic soft sections, mechanically stackable soft segments, fully decoupled rigid-link arms, reconfigurable cable robots with internally transforming end-effectors, and cable-driven parallel robots carrying serial manipulators (Islam et al., 2 Mar 2026, Luo et al., 18 Feb 2025, Zhang, 26 Oct 2025, Qi et al., 2019).

2. Mechanical architectures, routing, and reconfigurability

Mechanical architecture in FCRAs is dominated by how cable paths are routed and isolated. In the modular soft arm of (Islam et al., 2 Mar 2026), each segment contains a flexible internal backbone, embedded tendon-routing channels, and two end caps. The backbone is Clear PVC tubing with inner diameter $3\ \text{mm}$ and outer diameter $6\ \text{mm}$. A protective dual-helical tendon structure is formed by winding Kevlar threads around a $2\ \text{mm}$ metal rod mandrel in a dual helical pattern during casting; this provides tendon guidance and local reinforcement or protection.

In (Qi et al., 2024), routing is organized around section-level decoupling. Each actuation cable is fixed to an endcap, runs through embedded cable guides in a section, and then passes through a pathway in the endcap into the backbone tubing before being connected to the corresponding motor. Because the backbone length remains nearly constant during bending, the cable length changes of one section do not passively propagate to the others. This is a mechanical decoupling strategy rather than a control-layer approximation.

D3-Arm addresses routing by explicit low-friction decoupling mechanisms. Joint1 uses a decoupling cable aligner mechanism in which fixed pulleys attached to the base and movable pulleys attached to Link1 or Joint1 keep downstream cables aligned with the motion axis of Joint1. Joint2 and Joint3 use a decoupling rolling pair mechanism; the key relations are

$$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$

so the net cable length through the joint remains constant. A pretension force of 120 N is applied in the drive box to prevent slack and cable-off-pulley events (Luo et al., 18 Feb 2025).

Reconfigurability can also be moved into the end-effector. In (Zhang, 26 Oct 2025), an eight-cable robot uses an upper part, a lower part, a compression spring, a helical-grooved shaft, a matching spline nut, and a bearing. The end-effector has one relative translational and one relative rotational degree of freedom between its upper and lower components. Cable-induced linear compression is converted into rotation by the spring-driven screw-like coupling, and the bearing provides an additional rotational degree of freedom. The same principle is reused in a large-space cable-driven gripper; in that version, spring compression changes the gripper opening width, and the normally closed gripper retains holding force without continuous power.

General routing is itself an architectural design variable. The continuum robot of (Mahapatra et al., 2020) studies six routing patterns, including straight routing $(1,1,1,1,1,1,1,1,1,1)$, helical routing $(4,3,2,1,12,11,10,9,8,7)$, and four arbitrary routings. This demonstrates that cable path topology, not only actuator placement, shapes the admissible deformation of an FCRA.

3. Kinematics, statics, and workspace structure

A central modeling distinction in FCRAs is whether the arm is approximated geometrically or whether cable-body interaction is modeled explicitly. In the modular stackable soft arm of (Islam et al., 2 Mar 2026), workspace estimation uses a constant-curvature interpretation. The maximum radial reach is $R_{\max}$, the planar workspace area is

$A=\pi R_{\max}^2,$

and workspace volume is derived by integrating circular cross-sections along the arm’s axial direction. The single-segment trajectory was found to be near-circular, but the paper explicitly notes that under heavy payloads the actual shape departs from ideal constant curvature.

The nonlinear static model of (Qi et al., 2024) was developed precisely because constant-curvature geometry alone neglects transverse cable deformation. For a single section, the model introduces backbone curvature $\kappa_b$, cable curvature $\kappa_c$, and the cable “cutting into” the soft body through

$6\ \text{mm}$0

with the bending moment relation

$6\ \text{mm}$1

This yields a nonlinear actuation-to-curvature map that is solved numerically. For a multi-cable section, the model imposes

$6\ \text{mm}$2

The significance is not merely better curve fitting: the model is embedded into section-wise and whole-arm kinematics and then into inverse-kinematic motion planning (Qi et al., 2024).

For generally routed continuum robots, (Mahapatra et al., 2020) presents two complementary models. The first is a discrete optimization-based kinematic model that approximates each section as a virtual four-bar linkage and solves a constrained optimization problem with MATLAB fmincon. The second is a Cosserat rod static model with

$6\ \text{mm}$3

together with constitutive stiffness matrices $6\ \text{mm}$4 and $6\ \text{mm}$5. Both models matched experiment to within approximately 2%, while the optimization-based method was faster than the Cosserat method (Mahapatra et al., 2020).

Workspace arguments in FCRAs are often inseparable from transmission design. In (Zhang, 26 Oct 2025), a reconfigurable cable robot avoids the usual rotational restriction of fixed end-effectors by generating orientation internally; the end-effector can rotate through a wide range, even larger than 360 degrees, and geometric symmetry enables omnidirectional rotational capability. In (Pornthisan et al., 3 Jun 2026), the reduced 8-joint quaternion-joint arm had a smaller workspace than a prior 12-joint design when operated in $6\ \text{mm}$6, but achieved a comparable workspace when the joint bending limit was increased to $6\ \text{mm}$7. This suggests that joint count and per-joint bending range form an explicit workspace trade-off.

Hybrid cable-driven robots add elastic cable geometry directly into the generalized coordinates. In (Qi et al., 2019), the cable length is

$6\ \text{mm}$8

and cable elasticity is modeled as

$6\ \text{mm}$9

The corresponding structure matrix $2\ \text{mm}$0 maps platform twist to cable-length rates and cable tensions to platform wrench.

4. Control, redundancy, and motion planning

Control strategies in FCRAs range from purely kinematic regulation to hybrid model-based and data-driven schemes. The most explicit black-box formulation appears in (Liang et al., 21 Jul 2025), where the input is the motor-angle vector

$2\ \text{mm}$1

and the output is the joint-angle vector

$2\ \text{mm}$2

Using Hankel matrices and Willems’ fundamental lemma, the future output is written as

$2\ \text{mm}$3

and the resulting strictly convex quadratic program is solved using OSQP. A Data Selection Algorithm chooses the dataset with the smallest $2\ \text{mm}$4, reducing online solution time from about 19 ms to about 4 ms (Liang et al., 21 Jul 2025).

By contrast, (Islam et al., 2 Mar 2026) deliberately remains hardware-centric. Its experiments use open-loop actuation; no inverse kinematics is implemented, no curvature estimation is used for control, and motion is controlled empirically with visual observation and motion-capture feedback. This is a reminder that workspace demonstration and precision control are separable milestones in FCRA development.

Model-based inverse kinematics remains important when a tractable section model is available. In (Qi et al., 2024), the section variables are

$2\ \text{mm}$5

and motion planning uses a Jacobian pseudo-inverse,

$2\ \text{mm}$6

with a closed-loop correction term $2\ \text{mm}$7. In (Pornthisan et al., 3 Jun 2026), FABRIK provides a geometric inverse-kinematics baseline for quaternion-joint manipulators, while Residual Reinforcement Learning adds a learned residual correction using an asymmetric actor-critic architecture trained with PPO in Brax/MuJoCo.

Redundancy resolution is a recurring issue whenever cable count exceeds task-space dimension. The reconfigurable cable robot of (Zhang, 26 Oct 2025) addresses this structurally: the end-effector has six DoF, the design adds two additional relative DoF, the total becomes eight DoF, and the robot is driven by eight cables. Because the number of actuators matches the number of configuration DoF, the system is non-redundant and can be controlled purely through kinematics without tension sensing, torque control of motors, or sophisticated tension-distribution schemes. The converse appears in (Qi et al., 2019), where over-actuation requires

$2\ \text{mm}$8

so null-space tension can be redistributed to enforce positivity and optimize stiffness.

Safe motion planning must also include the cables themselves. In (Bury et al., 2019), continuous collision detection for a cable-driven parallel robot with an embarked robotic arm validates path segments by combining a lower bound $2\ \text{mm}$9 on inter-body distance with an upper bound $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$0 on relative velocity. A sufficient safety condition is

$$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$1

The method is integrated into the Humanoid Path Planner (HPP) and tested on CoGiRo.

5. Experimental performance and representative capabilities

Quantitative performance in FCRAs is highly architecture-dependent. The modular soft arm of (Islam et al., 2 Mar 2026) shows the strongest workspace-scaling result among the cited systems. Relative to the single-segment arm, the three-segment configuration achieved up to a 13-fold increase in planar workspace area and a 38.9-fold increase in workspace volume. The reported values were $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$2, $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$3, and $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$4 for one segment, versus $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$5, $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$6, and $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$7 for three segments. The same study quantified the material trade-off: for a 45 mm tendon pull with payloads from 0 to 200 g, Ecoflex 00-10 showed a bending angle drop from about 162° to 91°, a vertical tip displacement drop from about 110 mm to 57 mm, and a tendon tension increase from about 7.3 N to 9.8 N, whereas Ecoflex 00-50 required about 15–20 N tendon force and preserved geometric stability better under load (Islam et al., 2 Mar 2026).

D3-Arm emphasizes decoupled precision and dynamic capability. It demonstrated 1.29 mm average positioning error, 3.7855 mm 3-sigma repeatability, 2.0 kg payload capacity, maximum speed 1.47 m/s, and maximum acceleration 10.55 m/s². In direct decoupling verification, the maximum cable tension variation was 1 N, corresponding to a cable length change of less than 0.01 mm (Luo et al., 18 Feb 2025).

The data-driven MPC platform of (Liang et al., 21 Jul 2025) provides an alternative performance profile. On a real 3-segment FCRA with $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$8 motor inputs and $$\Delta l_1=\Delta l_{1,1}+\Delta l_{1,2}=\frac{R\theta_e}{2}, \qquad \Delta l_2=\Delta l_1=\frac{R\theta_e}{2},$$9 joint outputs at 50 Hz, the average positioning accuracy in five-point tests was approximately 2.070 mm. In five-point response tracking, the average tracking error was 1.418° for PID and 0.541° for data-driven MPC. The same platform also traced the letters “SMC” using quintic polynomial interpolation.

The modeling-centered soft arm of (Qi et al., 2024) shows that improving actuation-to-curvature fidelity translates into trajectory accuracy. For single-section endpoint tracking along a circular path, the average error was 1.58 cm for the proposed model versus 2.43 cm for the baseline, a 35% reduction. For a two-section arm, circular trajectory tracking was 3.76 cm versus 5.92 cm, a 36% reduction, and straight trajectory tracking was 1.70 cm versus 3.52 cm, a 52% reduction (Qi et al., 2024).

In generally routed continuum robots, the experimental backbone-shape error stayed below 3.6 mm, corresponding to less than 2% of the total robot length, under 400 g loads. The optimization-based model took about 2.5 s on average, versus about 10.5 s for the Cosserat rod model (Mahapatra et al., 2020). The same paper also demonstrated a three-fingered gripper gripping a spherical object, gripping a cube, and manipulating a cube.

For reconfigurable cable robots, (Zhang, 26 Oct 2025) reports that internally generated rotation can exceed 360 degrees and that the same mechanism can support a large-space cable-driven gripper. For quaternion-joint redundant manipulators, (Pornthisan et al., 3 Jun 2026) reports that for 1,000 random poses, a prior baseline achieved 77.6% success in 0.3705 s, while the proposed FABRIK method achieved 70.4% success in 0.0221 s. The abstract and conclusion state that Residual Reinforcement Learning outperforms FABRIK by approximately three orders of magnitude in positional and orientational accuracy, while the details also state that the base controller alone can achieve a median position error of 3.7 mm and sub-degree orientation error and that the residual policy after 50M steps had not yet reached deployment-ready performance (Pornthisan et al., 3 Jun 2026).

6. Limitations, recurring misconceptions, and research directions

Several limitations recur across otherwise different FCRA embodiments. Cable elasticity, resilience, hysteresis, and friction remain fundamental modeling obstacles (Liang et al., 21 Jul 2025). In soft modular arms, the constant-curvature model can break down under load, self-weight and tendon loading accumulate as segments are added, and scalability has a practical limit unless actuation and structural support are improved (Islam et al., 2 Mar 2026). In D3-Arm, cable elasticity still limits stiffness and accuracy, payload capacity is moderate rather than high, and cable detachment becomes an issue at heavier loads (Luo et al., 18 Feb 2025).

A second recurring limitation concerns the gap between kinematic feasibility and physically robust execution. The non-redundant cable robot of (Zhang, 26 Oct 2025) can be controlled purely kinematically because its eight cables match eight configuration DoF, but the paper still states that motion should be realized by commanding cable lengths, ensuring proper tensioning, and optimizing tension magnitude with respect to safety and energy criteria. In hybrid systems, (Bury et al., 2019) explicitly notes that continuous collision validation does not guarantee that a path is physically executable on the real robot; it does not by itself optimize cable tensions, dynamic feasibility, comfort margins, or maximal clearance.

A third misconception is that adding cables or joints monotonically improves performance. The reconfigurable cable robot literature states that increasing the number of cables enlarges translational workspace but introduces cable interference and non-unique tension solutions (Zhang, 26 Oct 2025). The quaternion-joint manipulator literature shows the dual statement: reducing joint count can preserve workspace if each joint is allowed a larger bending range, but the control model becomes more challenging and learning-based correction remains nontrivial (Pornthisan et al., 3 Jun 2026).

The cited future directions are correspondingly specific. They include closed-loop control, better kinematic modeling, mitigation of self-weight effects in longer modular chains, more sophisticated grippers, systematic modeling, workspace analysis, stiffness evaluation, parameter optimization, motion planning, feedback control, grasping, improved adaptability and robustness in dynamic environments, hybrid position-force control for grinding, polishing, and cleaning, virtual reality motion platforms, and underwater grasping with actuators above the water (Islam et al., 2 Mar 2026, Liang et al., 21 Jul 2025, Zhang, 26 Oct 2025). Taken together, these works suggest that the contemporary FCRA field is defined less by a single morphology than by a common technical program: remote cable actuation, mechanical or algorithmic decoupling, explicit treatment of cable-induced nonlinearity, and workspace expansion without sacrificing controllability.