Soft Continuum Arms (SCAs)

• Soft continuum arms (SCAs) are compliant manipulators defined by continuously distributed deformations, where curvature, torsion, and strain fields describe their state.
• Their embodiments range from fiber-reinforced pneumatic devices and tendon-driven elastomeric rods to wearable fabric extra limbs, highlighting diverse designs for safe human interaction.
• Advanced control and sensing methods, such as Cosserat-rod modeling, visual servoing, and reinforcement learning, address challenges in modeling infinite-dimensional deformation and achieving precise manipulation.

Soft continuum arms (SCAs) are compliant manipulators whose configuration changes continuously along the body, so their state is more naturally expressed through curvature, torsion, extension, shear, or strain fields than through a small set of rigid-joint coordinates. Across the literature, SCAs appear as pneumatically actuated fiber-reinforced arms, tendon-driven elastomeric rods, wearable fabric extra limbs, hybrid soft-rigid continuum devices, and even continuum morphing aerial arms. Their appeal is tied to compliance, safe and adaptive interaction, conformal contact, and access to bending, twisting, extension, and wrapping behaviors in cluttered or human-facing environments; their central difficulty is that continuous deformation, high compliance, and effective infinite-dimensionality make modeling, sensing, and control inseparable system-level problems rather than isolated subproblems (AlBeladi et al., 2020, Nguyen et al., 2019, Yang et al., 23 Apr 2025, Perfetta et al., 17 Mar 2026).

1. Embodiments and structural archetypes

SCAs are not a single hardware class but a family of embodiments defined by distributed deformation. The BR2 platform is a single-section pneumatic soft continuum manipulator built from a parallel arrangement of Fiber Reinforced Elastomeric Enclosures (FREEs), with one soft bending actuator and two soft rotating actuators, so it can bend and rotate simultaneously in 3D; in experimental studies it appears both as a twisting-and-bending perception benchmark and as a platform for visual servoing and sim-to-real reinforcement learning (Yang et al., 23 Apr 2025, AlBeladi et al., 2020, Kamtikar et al., 2022). SoPrA is a two-segment pneumatically actuated, fiber-reinforced soft continuum arm with three chambers per segment and integrated capacitive flex sensors, explicitly designed so that analytical modeling and proprioceptive state estimation can be developed together (Toshimitsu et al., 2021). Other architectures are more specialized: the fabric Soft Poly-Limb is a wearable, three-segment, fabric-based continuum extra limb designed for activities of daily living (Nguyen et al., 2019); a single-section inextensible continuum arm combines three pneumatic muscle actuators with a hyper-redundant rigid-link backbone (Nazari et al., 2019); and a prismatic soft actuator has been mounted beneath a two-segment SoPrA arm to add a translational degree of freedom at the base (Wand et al., 2022).

This diversity extends beyond classical manipulation. QuadSoft uses four tendon-driven, continuous-curvature soft robotic arms as quadrotor appendages, with a semi-rigid center and flexible arms that bend to redirect thrust without changing the basic rotor topology (Verdin et al., 1 Apr 2026). At the other extreme, a 38 cm conical arm cast from Ecoflex 00-50 preserves distributed bending, torsion, and limited axial compression without hardware discretization or stiffness-based mode suppression, and is used to argue that a fully compliant continuum body need not be reduced to a quasi-rigid serial chain in order to admit high-performance control (Perfetta et al., 17 Mar 2026).

The mechanical choices across these systems reveal several recurrent design logics. Fiber reinforcement is used to suppress radial ballooning and concentrate deformation into useful axial or bending modes (Toshimitsu et al., 2021, Doroudchi et al., 2022). Tapered or anisotropic cross-sections allocate compliance and force capacity nonuniformly along the arm (Toshimitsu et al., 2021, Verdin et al., 1 Apr 2026). Hybridization with rigid guides or backbones appears when lateral stability, inextensibility, or precise routing is required, as in the prismatic base module and the inextensible hybrid PMA arm (Wand et al., 2022, Nazari et al., 2019). This suggests that the continuum designation in SCA research refers less to complete material softness than to the presence of spatially distributed deformation and the need to reason about shape as a field.

2. Geometric and mechanical descriptions

A dominant mathematical language for SCAs is Cosserat-rod or geometric-mechanics modeling on $SE(3)$. In one common formulation, the arm configuration along arc length $s\in[0,L]$ is written as

$$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$

where $\mathbf{q}(s)\in\mathbb{R}^3$ contains stretching/shearing strains and $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$ contains bending/twisting strains (AlBeladi et al., 2020). A closely related posture-reconstruction formulation uses

$$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$

with $\xi(s)\in\mathfrak{se}(3)$ containing bending, twist, shear, and axial stretch channels (Wang et al., 2024). Under negligible shear and stretching, BR2 is also modeled with constant curvature $\kappa$ and constant torsion $\tau$, using

$r'(s)=R(s)\,v(s),\qquad R'(s)=R(s)\,\hat u(s),$

with $s\in[0,L]$0 and $s\in[0,L]$1 (Yang et al., 23 Apr 2025).

Reduced-order parameterizations differ mainly in how they compress the strain field. Piecewise constant curvature (PCC) remains a common engineering approximation, especially for dynamic control and operational-space formulations (Doroudchi et al., 2022, Fischer et al., 2022, Kazemipour et al., 2021). More general strain-basis approaches expand $s\in[0,L]$2 or $s\in[0,L]$3 in chosen basis functions, so constant curvature, $s\in[0,L]$4-segment piecewise constant curvature, linear curvature, quadratic curvature, cubic curvature, and PCA-derived strain bases all appear as special cases of the same reduced-order viewpoint (AlBeladi et al., 2020, Wang et al., 2024). A variable-strain formulation goes further by approximating the full six-dimensional Cosserat strain field as

$s\in[0,L]$5

thereby retaining nonuniform distributed strain without collapsing the arm into a small set of constant-curvature sections (Perfetta et al., 17 Mar 2026).

A distinct modeling subclass imposes inextensibility explicitly. For the single-section inextensible continuum arm with three PMAs and a rigid-linked backbone, the centerline arc length remains fixed,

$s\in[0,L]$6

and the actuator-length variations satisfy

$s\in[0,L]$7

This removes the usual extension degree of freedom of a three-PMA segment and yields a two-dimensional reduced-order shape map in $s\in[0,L]$8 for a constant-length circular arc (Nazari et al., 2019). By contrast, antagonistic design analysis treats the arm as a planar Cosserat-like structure under load and asks whether a required internal wrench sequence lies inside the set generated by feasible pressure inputs, turning embodiment comparison into a static mechanics problem rather than a purely kinematic one (Fan et al., 2024).

These models differ in fidelity, computational cost, and what they preserve. PCC simplifies geometry and control but suppresses some distributed modes. Constant curvature/constant torsion is efficient for RL rollout and visual servoing but introduces mismatch at high torsion or under strong loading (Yang et al., 23 Apr 2025). Variable-strain and full Cosserat formulations retain richer mechanics and are more suitable when bending, twist, axial compression, or underactuation cannot be ignored (Wang et al., 2024, Perfetta et al., 17 Mar 2026, Doroudchi et al., 2022). A persistent theme is that reduced-order modeling in SCA research is not optional; it is the practical mechanism by which infinite-dimensional shape is made observable, controllable, or optimizable.

3. Sensing, proprioception, and shape reconstruction

For SCAs, shape sensing is not an accessory but the state-estimation problem itself. Because these robots deform continuously, classical rigid-joint sensing such as encoders is not applicable, and accurate shape knowledge is needed for end-effector positioning, obstacle avoidance, contact reasoning, model-based control, and autonomy (AlBeladi et al., 2020). One line of work uses exteroception. A monocular base-mounted fisheye camera can be combined with a geometric strain parameterization of the arm and a nonlinear least-squares fit between observed and predicted left/right image boundaries. On the BR2 arm, five basis families were compared—constant, piecewise constant, linear, quadratic, and cubic—and the best overall result came from a two-segment piecewise constant strain model, with

$s\in[0,L]$9

over the tested workspace, and in the practically relevant upper-half “Region A”

$$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$0

so the mean tip error there is less than the arm radius of $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$1 (AlBeladi et al., 2020).

A second exteroceptive line replaces optimization with learned strain inference. A neural network can take sparse noisy marker-pose measurements and predict finite-dimensional strain coefficients, which are then integrated through Cosserat kinematics to reconstruct a smooth posture. On a simulated octopus arm and a physical BR2 manipulator, this framework was reported to be over five orders of magnitude faster than a forward-backward optimization baseline while maintaining comparable accuracy; on the octopus arm, average reconstruction errors across 27 targets were reported to be below $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$2 (Wang et al., 2024). This suggests that for some SCAs, the critical computation is not dense shape fitting itself but the inference of a low-dimensional strain representation from sparse measurements.

Embedded sensing provides a different route. SoPrA integrates one two-axis capacitive flex sensor per segment and uses the linear relation $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$3 as a constraint in a quadratic program that corrects model-predicted shape and estimates external tip loads without relying on external motion capture during estimation (Toshimitsu et al., 2021). The same work reports that absolute tip-force estimation is poor but relative load estimation is approximately linear and consistent across poses, which is often the practically useful part for contact-aware behavior. By contrast, ManiSoft deliberately withholds internal soft-body states from benchmarked policies in order to mirror a recurring empirical fact: soft arms often have unreliable or absent proprioception, and visual estimation of the deformed body becomes a bottleneck for control (Wei et al., 18 May 2026).

Several failure modes recur across sensing papers. Monocular reconstruction degrades near full extension, where image observations become ill-conditioned, and near the edge of a fisheye image, where distortion and calibration errors matter more (AlBeladi et al., 2020). Learned reconstruction inherits the coverage limits of the simulator-generated training set and remains arm-specific (Wang et al., 2024). Benchmark results in ManiSoft attribute many failures to inaccurate visual estimation of proprioceptive state and to policies that do not exploit deformation intentionally (Wei et al., 18 May 2026). A consistent implication is that SCA sensing is not merely about observing the tip: it is about inferring a distributed latent shape with enough fidelity for control and interaction.

4. Control, visual servoing, and dynamic regulation

Control approaches for SCAs span model-based inverse dynamics, operational-space control, visual servoing, and reinforcement learning, but they differ mainly in where they place model complexity. A Cosserat-rod-based controller for a multi-segment silicone arm demonstrates closed-loop tracking of configuration-space variables using real-time simulation to estimate unmeasured internal variables from segment-tip poses measured by motion capture; the controller reshapes a two-segment arm into bending and extension configurations in 3D and is presented as the first experimental demonstration of closed-loop control of soft-arm configuration-space variables for an independently controllable multi-segment arm (Doroudchi et al., 2022). This is a direct continuum-mechanics approach: measurement is sparse, hidden variables are simulated, and inverse dynamics are solved online.

A related but more explicitly robust line uses adaptive sliding mode control on a two-segment pneumatic soft continuum manipulator. Its PCC-based Euler–Lagrange model uses centroid-based segment inertia rather than a tip-lumped approximation, and the adaptive controller combines a regressor-based dynamic compensation term with an adaptive disturbance-bound estimate. In experiments under different payloads, the tracking performance is reported to be around $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$4 more accurate than that of the inverse dynamics baseline, and the centroid-based model avoids the roughly $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$5 steady-state position error seen in the tip-lumped alternative (Kazemipour et al., 2021). The same paper argues that the approach can be generalized to any continuum robotic arm with an arbitrary number of segments.

Operational-space control extends this logic from curvature coordinates to task coordinates. On SoPrA, dynamic task-space control is reported to improve tracking accuracy by $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$6 and increase speed by a factor of $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$7 compared to the state of the art for task-space control, while enabling pick-and-place, obstacle avoidance, throwing objects with a soft gripper, and deliberate force application by drawing with chalk (Fischer et al., 2022). These results matter because they directly target a common SCA weakness: quasi-static task-space methods are accurate only when motions are slow and external forces are small.

Visual servoing shifts complexity from shape mechanics to image-space inference. In a structured environment, a distal RGB camera on BR2 can be used with a CNN that predicts actuation directly from the current image, together with a proportional update law in predicted-actuation space. The integrated image-to-actuation approach achieves average Euclidean position error $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$8 and average rotation error $$\mathbf{X}(s)= \begin{bmatrix} \mathbf{R}(s) & \mathbf{p}(s)\ \mathbf{0} & 1 \end{bmatrix}\in SE(3), \qquad \frac{\partial \mathbf{X}(s)}{\partial s}=\mathbf{X}(s)\mathbf{\Omega}(s),$$9 over 30 random targets, outperforming a modular image-to-pose-to-actuation approach, and remains robust to new targets, lighting changes, distributed load, and partial loss of bending functionality (Kamtikar et al., 2022). In a more recent learning-based visual-servoing framework, RL is trained entirely in Gazebo under a constant-curvature and constant-torsion model and deployed on BR2 with no hardware fine-tuning. The controller operates in $\mathbf{q}(s)\in\mathbb{R}^3$0 space, while a local controller on hardware realizes the commanded configuration; the result is a reported $\mathbf{q}(s)\in\mathbb{R}^3$1 success rate in simulation and $\mathbf{q}(s)\in\mathbb{R}^3$2 hardware success in zero-shot sim-to-real transfer for centering a target in the distal camera (Yang et al., 23 Apr 2025).

A common misconception is that all learning-based SCA control is end-to-end actuation learning. The zero-shot framework argues the opposite: the transferable part is the configuration-space geometry, not the actuation-to-configuration map, so the RL controller is trained only on the kinematic layer and a local controller handles the hardware-specific correction (Yang et al., 23 Apr 2025). Conversely, highly model-based approaches show that even sparse sensing can suffice when internal state is reconstructed online through continuum dynamics (Doroudchi et al., 2022). The control literature therefore divides less by “model-based versus learned” than by where the arm’s complexity is absorbed: in mechanics, in perception, or in the decomposition between the two.

5. Workspace, manipulation, and grasping

A recurring practical limitation of many pneumatically actuated SCAs is workspace shape. In the common coaxial multi-segment architecture, adding more bending sections does not substantially alter the fact that the reachable set remains approximately hemispherical. A base-mounted prismatic soft actuator changes that geometry by translating the entire arm vertically, increasing fitted workspace volume by $\mathbf{q}(s)\in\mathbb{R}^3$3, or a factor of $\mathbf{q}(s)\in\mathbb{R}^3$4, and enabling motions such as straight lines that were previously infeasible (Wand et al., 2022). On the augmented SoPrA system, average tip tracking errors were reported as $\mathbf{q}(s)\in\mathbb{R}^3$5 for a helix, $\mathbf{q}(s)\in\mathbb{R}^3$6 for an inclined circle, and $\mathbf{q}(s)\in\mathbb{R}^3$7 for a line, while tabletop pick-and-place and container-access tasks were reproducible with a $\mathbf{q}(s)\in\mathbb{R}^3$8 success rate over 10 runs each (Wand et al., 2022).

Manipulation capability also depends strongly on whether the arm exploits distributed body contact rather than only distal tip forces. The wearable fabric Soft Poly-Limb illustrates this continuum advantage clearly. Its three fabric segments produce an experimentally measured workspace volume of $\mathbf{q}(s)\in\mathbb{R}^3$9, maximum vertical range $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$0, and maximum horizontal range $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$1. In free-space loading, the $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$2 arm lifts $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$3 while outstretched and parallel to the ground, but when allowed to wrap around and support an object against the wearer’s body, it carries $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$4, equal to $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$5 its own weight (Nguyen et al., 2019). This distinction between cantilever payload and whole-body wrapped payload is central to SCA manipulation: the useful wrench transmission of a continuum body is often tied to distributed contact geometry rather than tip force alone.

Planar grasping theory formalizes this geometric viewpoint. If the object boundary is treated as the arm’s “shadow curve,” the relative state can be reduced to distance $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$6 and contact angle $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$7, with the arm curvature $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$8 as control: $\boldsymbol{\kappa}(s)\in\mathbb{R}^3$9 An optimal control problem then synthesizes grasping shapes, while distributed contact is summarized through a continuum grasp map

$$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$0

and grasp-quality metrics $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$1, $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$2, and $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$3 derived from the Gramian $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$4 (Halder et al., 10 Apr 2026). This reframes continuum grasping as a boundary-following and wrench-transmission problem rather than as a discrete point-contact problem.

Task competence under load can also be evaluated before fabrication. For antagonistic, pneumatically driven continuum arms, a wrench-space method tests whether the required internal wrench sequence for a loaded target shape lies inside the set generated by feasible pressures. The method introduces absolute and relative unattainability metrics, produces graphical segment-wise wrench interpretations, and is reported to be over $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$5 faster than optimization-based methods; in the demonstrated comparison, computation took $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$6 s versus $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$7 s, an $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$8 speed increase (Fan et al., 2024). One practical consequence is that antagonistic architectures can be compared against non-antagonistic ones on task-specific load-bearing capability, rather than by workspace alone.

6. Benchmarks, misconceptions, and open directions

Several recurrent assumptions in SCA research are now being challenged. One is that richer reduced-order bases always improve sensing or control. On BR2, increasing polynomial order beyond quadratic did not improve monocular reconstruction, and the best visual shape estimator used a two-segment piecewise constant strain basis rather than higher-order polynomial bases (AlBeladi et al., 2020). Another is that simply adding more coaxial bending segments solves practical manipulation. The prismatic-base study argues that additional translational degrees of freedom change workspace shape in a way more segments do not (Wand et al., 2022). A third is that distributed softness and dynamic precision are incompatible. A fully compliant, non-segmented continuum arm with direct-drive tendon actuation reports a task-accomplishment velocity of $$q(s)= \begin{bmatrix} Q(s) & x(s)\ 0 & 1 \end{bmatrix}\in SE(3), \qquad \partial_s q(s)=q(s)\,\xi(s),$$9 at $\xi(s)\in\mathfrak{se}(3)$0 accuracy and $\xi(s)\in\mathfrak{se}(3)$1 at $\xi(s)\in\mathfrak{se}(3)$2 accuracy, versus a prior best listed value of $\xi(s)\in\mathfrak{se}(3)$3 for millimetric precision, and a peak tip speed of $\xi(s)\in\mathfrak{se}(3)$4 in a pendulum-striking task (Perfetta et al., 17 Mar 2026). The paper presents this as evidence that a fully compliant continuum body can achieve fast, precise task-space regulation without being mechanically reduced to piecewise quasi-rigid modules.

Benchmarking work has also made the field’s bottlenecks more explicit. ManiSoft defines four language-conditioned soft-arm tasks, provides $\xi(s)\in\mathfrak{se}(3)$5 scene–trajectory pairs, and shows that standard vision-language manipulation models perform modestly even in clean scenes and often collapse under randomization: average success is $\xi(s)\in\mathfrak{se}(3)$6 for Diffusion Policy, $\xi(s)\in\mathfrak{se}(3)$7 for RDT, and $\xi(s)\in\mathfrak{se}(3)$8 for OpenVLA-OFT in clean scenes, versus $\xi(s)\in\mathfrak{se}(3)$9, $\kappa$0, and $\kappa$1 under randomized conditions (Wei et al., 18 May 2026). The failure analysis attributes errors primarily to inaccurate visual estimation of proprioceptive state and to policies that do not intentionally use deformability for obstacle avoidance. This directly contradicts the informal assumption that compliance will automatically help high-level policies once the arm is made soft.

Beyond manipulation, SCA dynamics have also been examined as computational resources. A three-section pneumatic continuum arm used as a physical reservoir shows information-processing capability that varies with input bandwidth and amplitude, with optimal regimes for tasks such as NARMA emulation and nonlinear memory capacity (Torres et al., 2018). This suggests a broader interpretation of continuum dynamics: the body is not only a plant to be controlled, but also a distributed dynamical system that can perform temporal filtering and nonlinear transformation. A plausible implication is that future SCA controllers may exploit morphology as part of the computation, rather than treating all deformation as disturbance.

Open problems identified across the literature are consistent. Vision-based reconstruction still needs automated image extraction, robustness to occlusion, and better treatment of calibration drift (AlBeladi et al., 2020). Fast learned reconstruction remains simulator- and arm-specific (Wang et al., 2024). Dynamic controllers frequently depend on external motion capture or simplified actuator models (Toshimitsu et al., 2021, Doroudchi et al., 2022, Kazemipour et al., 2021). Sim-to-real visual servoing still suffers from model mismatch, lack of depth information, and reduced precision in small corrective motions (Yang et al., 23 Apr 2025). Benchmarks expose a gap between rigid-arm multimodal reasoning and deformable-body control (Wei et al., 18 May 2026). Taken together, the field is moving from isolated demonstrations of soft-arm compliance toward integrated systems in which embodiment, reduced-order mechanics, proprioception, exteroception, and task-level reasoning are designed jointly rather than sequentially.