Reconfigurable Tendon-Driven Robots

• RTR is a tendon-driven robot with reconfigurable routing, stiffness, and morphology that adapts its actuation-to-shape mapping for varied tasks.
• These systems use modular designs and variable stiffness techniques to optimize payload, workspace, and dynamic control in applications like minimally invasive surgery and locomotion.
• RTR architectures balance compliance and stability through innovative mechanics and learning-based control, addressing challenges in high-dimensional actuation spaces.

Searching arXiv for the cited RTR-related papers to ground the article in current literature. A Reconfigurable Tendon-Driven Robot (RTR) is a tendon-driven robot in which tendon routing, stiffness distribution, morphology, active topology, or controller-facing shape variables are deliberately reconfigured to alter deformation modes, workspace, payload, locomotion, or environment interaction. In the cited literature, RTRs appear as modular bead-based continuum manipulators with segment-wise variable stiffness and internal routing, serial compliant snake robots whose globally routed tendons switch the body between vertical bending and axial twisting modes, tendon-driven continuum manipulators with actively rotating spacer disks that reroute tendons during operation, independently lockable-joint bodies that eliminate inter-segmental coupling, origami-inspired manipulators whose joint stiffnesses are optimized alongside tendon commands, separable medical manipulators whose reusable and disposable sections are reconfigured around a long passive proximal section, and soft quadrupeds whose tendon-driven morphology is used as a control abstraction rather than as a fixed joint-space plant (Walker et al., 2024, Kwon et al., 2024, Dash et al., 14 Apr 2026, Lin et al., 23 Jul 2025, Chen et al., 14 Apr 2025, DeBuys et al., 2023, Niu et al., 2024).

1. Conceptual scope and defining characteristics

Within this literature, RTR denotes more than a robot that is merely “tendon actuated.” The narrower idea is that routing is not fixed and can be deliberately changed to alter deformation modes, or that the robot’s geometry and stiffness can be reconfigured in operation, or that selected joints can be temporarily converted into rigid links so that the active kinematic chain changes step by step (Dash et al., 14 Apr 2026, Walker et al., 2024, Lin et al., 23 Jul 2025). This distinguishes RTRs from conventional fixed-routing tendon-driven continuum manipulators, in which tendon paths are predetermined by spacer-disk holes or by a fixed outer circumference.

The reconfiguration variables reported in the cited work span several layers of embodiment. At the structural level, modular length and assembly allow bead or segment addition and removal, and a reusable/disposable split allows the actuation body and distal catheter to be reconfigured around different passive proximal shapes (Walker et al., 2024, Ouyang, 2019, DeBuys et al., 2023). At the actuation-topology level, globally routed tendons on the back, abdomen, or spiral body surface switch a snake robot among upper vertical bending, lower vertical bending, and axial twisting modes, while rotating spacer disks alter local tendon routing angles during or after actuation (Kwon et al., 2024, Dash et al., 14 Apr 2026). At the stiffness level, tendon-induced compressive jamming, layer jamming, Joule-heated variable stiffness joints, and tapered geometry all change the mapping from tendon input to backbone shape (Walker et al., 2024, Ouyang, 2019, Chen et al., 14 Apr 2025, Hansen et al., 19 Mar 2026).

A common misconception is that reconfigurability in tendon-driven robots is only geometric. The cited systems show that RTRs can be reconfigured through stiffness, tendon path, locking state, proximal configuration, or control-space parameterization. A plausible implication is that RTR is best treated as a systems category defined by deliberate alteration of the actuation-to-shape map, rather than by any single mechanism.

2. Representative architectures and embodiments

The reported architectures range from discrete-continuum manipulators to locomoting bodies and catheter-scale devices. The modular bead-style continuum manipulator in "A Modular, Tendon Driven Variable Stiffness Manipulator with Internal Routing for Improved Stability and Increased Payload Capacity" consists of 20 rigid beads connected by flexible TPU hinges, with each consecutive pair arranged at a 90° offset to produce orthogonal planar motion and, collectively, spatial mobility. The prototype is approximately 400 mm long in the flexible state; each bead is 26 mm long, 38 mm wide, and about 7 g, while each hinge is about 19 mm high and less than 1 g. The beads are FDM-printed PLA with 15% infill, and the hinges are FDM-printed TPU (Shore 95A) with 15% infill (Walker et al., 2024).

The snake-robot instantiation in "Development of Tendon-Driven Compliant Snake Robot with Global Bending and Twisting Actuation" is a 13-module serial body with 12 continuum joints, about 1.25 m long and weighing 2.21 kg. Its actuation layout is 10 single-axis motors plus 3 dual-axis motors. Each continuum joint uses a coil spring plus a central backbone, and the backbone includes a torque coil with a NiTi rod and PTFE/PU tubes inside to improve torsional behavior (Kwon et al., 2024).

The actively reroutable continuum manipulator in "Actuation space reduction to facilitate insightful shape matching in a novel reconfigurable tendon driven continuum manipulator" is 560 mm long with eight equal-length segments. Each spacer disk can rotate about the backbone axis using a servo motor, bearing, and spur gear, and the tendon hole is located at a radius of 34 mm from the backbone centerline. Rotating a disk changes the local tendon routing angle between the disks above and below it (Dash et al., 14 Apr 2026).

Other RTR embodiments broaden the category further. The origami-inspired manipulator uses multiple serially connected modules with top and bottom triangular plates, three pairs of vertical and diagonal links, and silicone-tube compliant spherical joints; it is actuated by three zigzag-routed tendons and reconfigured through variable stiffness joints (Chen et al., 14 Apr 2025). The separable medical TDRM divides the system into a reusable back portion containing four Faulhaber linear motors and a disposable front portion containing the tendons, long passive proximal section, and articulation section, with a spool ratio $R:r = 3:1$ in the prototype (DeBuys et al., 2023). The tapered-polymer continuum robot uses a TPU backbone, rigid discs, three equally spaced tendons, and an integrated cylindrical electronics housing with Dynamixel XH430-210T motors and FX29 load cells; the validation robot had a 34.5 cm backbone, 10 discs, base radius 1.11 cm, tip radius 0.45 cm, and taper angle about $1.08^\circ$ (Hansen et al., 19 Mar 2026). The lockable-joint RTR employs a seven-joint body with a six-motor actuation module, of which four motors handle bending and two motors handle locking and unlocking selection and switching (Lin et al., 23 Jul 2025).

Embodiment	Reconfigurable element	Representative feature
Modular bead continuum manipulator (Walker et al., 2024)	Length, section stiffness, tendon routing	Two PCC segments, internal routing, 1 kg payload result
Compliant snake robot (Kwon et al., 2024)	Global tendon-routing mode	Top, bottom, and spiral routing for three locomotion principles
Rotating-disk RTDCM (Dash et al., 14 Apr 2026)	Active spacer-disk rotation	Routing can change prior to or after actuation
Origami-inspired manipulator (Chen et al., 14 Apr 2025)	Joint stiffness vector $S$	Stiffness-policy co-optimization
Separable catheter TDRM (DeBuys et al., 2023)	Reusable/disposable split, proximal configuration adaptation	Four-tendon open-loop redundant control
Lockable-joint RTR (Lin et al., 23 Jul 2025)	Joint locked/free state	Coupling elimination with only six motors
Tapered TPU continuum robot (Hansen et al., 19 Mar 2026)	Parametric geometry and graded stiffness	Taper-aware Cosserat model and tension sensing

3. Reconfiguration modalities and mechanical principles

One major reconfiguration mode is modular assembly. The bead-based manipulator can change length by adding or removing bead modules, and the 2019 layer-jamming continuum robot uses segment-to-segment connectors with tapered holes, protruded cones, and screw-fastened halves so that arm length is not fixed and new modules can be added in principle (Walker et al., 2024, Ouyang, 2019). In both cases, the discrete geometry approximates a continuum-like shape while remaining serviceable, repairable, and scalable.

A second mode is variable stiffness. In the bead-based manipulator, stiffness is achieved by applying a tensile load of 30 N to each tendon through the motors, compressing the bead stack, increasing frictional contact between beads and hinge interfaces, and producing a stiffer state through compressive jamming and frictional locking. The paper further notes that out-of-plane jamming creates contact between both the top and side of the bead rails, increasing shear contact and thus friction (Walker et al., 2024). In the 2019 modular continuum robot, variable stiffness is produced by vacuum-induced layer jamming: at approximately atmospheric pressure the internal layers can slip, whereas vacuum increases interlayer friction and locks the segment (Ouyang, 2019). In the origami-inspired manipulator, variable stiffness is realized by Joule heating of variable stiffness joints, making the stiffness vector $S$ itself a reconfiguration parameter (Chen et al., 14 Apr 2025).

A third mode is routing reconfiguration. The snake robot uses globally routed tendons fixed at the head and tail and wound by the second axes of the middle dual-axis motors. Straight routing along the back/top surface yields upper vertical bending, straight routing along the abdomen/bottom surface yields lower vertical bending, and spiral routing around the body yields axial twisting (Kwon et al., 2024). The rotating-disk RTDCM makes routing a direct input: $$\text{actuation space} = \{\text{tendon displacement}\} \cup \{\text{disk rotation angles}\}.$$ Each disk rotation adds a degree of freedom and changes the local routing geometry between adjacent disks (Dash et al., 14 Apr 2026). The modular bead manipulator uses a different routing strategy: tendons pass through the center of sections that should remain stable and transition to the outside only for the segment being actuated, thereby decoupling tangential tension from unwanted proximal bending moment (Walker et al., 2024).

A fourth mode is topological reconfiguration through locking. In "Reconfigurable Tendon-Driven Robots: Eliminating Inter-segmental Coupling via Independently Lockable Joints," each joint can be switched between free and locked states by a pair of antagonistic tendons acting on an asymmetric trigger and slider. The dead-point geometry means the state is mechanically maintained without continuous power supply (Lin et al., 23 Jul 2025). This is a distinct mechanical principle: reconfiguration is not only a change of shape or stiffness, but a change in which joints belong to the active chain.

A fifth mode is geometry reconfiguration by parameterized design. The tapered TPU continuum robot treats backbone length, taper angle, base diameter, disc count, and cable-hole diameter as user-specified parameters in Inventor iLogic, so the geometry and stiffness distribution can be regenerated and fabricated for different compliance-workspace tradeoffs (Hansen et al., 19 Mar 2026). This suggests a design-time form of RTR distinct from in-operation rerouting or locking.

4. Modeling, inverse problems, and control abstractions

RTR modeling spans kinematic approximations, energy-based mechanics, Cosserat-rod formulations, and learning-based surrogates. The bead-based manipulator is modeled as two independently actuated segments under Piecewise Constant Curvature (PCC), with segment state

$q_i = [\phi_i, \theta_i]^T \in \mathbb{R}^2,$

where $\phi_i$ is the bending plane angle and $\theta_i$ is the curvature angle. The model assumes no twist, no elongation, minimal stiffening-induced length change below 2%, and structural interfaces that resist twisting; OptiTrack motion capture showed good agreement between recorded and modeled planar tip positions (Walker et al., 2024).

The actively reroutable TDCM emphasizes the inverse problem created by dynamic routing. Desired backbone shape must be mapped not only to tendon displacement but also to disk selection, rotation direction, and rotation angle. To reduce this high-dimensional nonlinear actuation space, the paper projects backbone shape into curvature–torsion space and exploits the empirical observation that torsion sign changes indicate disk rotation locations, torsion sign indicates rotation direction, and curvature correlates with tendon pull magnitude. This leads to a four-step sequential shape-matching strategy: infer disk indices and directions from torsion, choose tendon displacement by minimizing curvature RMSE using Golden Section search, refine disk angles for global shape RMSE, and finally rotate the eighth disk within $[-20^\circ,20^\circ]$ to reduce tip error (Dash et al., 14 Apr 2026).

Energy-based mechanics appear in the origami-inspired manipulator, where shape is predicted by minimizing total potential energy subject to tendon-length and force-limit constraints. The design variables are the VSJ stiffnesses, represented by the vector $S$, and the control variables are the tendon displacements $1.08^\circ$0. The controller is trained with PPO, and the joint optimization treats the design distribution $1.08^\circ$1 over stiffness parameters together with the policy $1.08^\circ$2 (Chen et al., 14 Apr 2025). This makes reconfiguration parameters explicit in the control objective rather than treating them as fixed background constants.

Two strands of literature move from mechanics toward learned control. In SoftQ, each leg pose is parameterized by bending angle $1.08^\circ$3, rotational angle $1.08^\circ$4, and compression length $1.08^\circ$5, so gait optimization acts in a morphology-aware space rather than through 12 raw motor commands. The model-based reinforcement learning pipeline restricts the action to four gait parameters for diagonal leg pairs, trains a DNN surrogate on simulator data, uses SAC for policy learning, and performs post-training in the high-fidelity simulator (Niu et al., 2024). In tendon-force-aware sim-to-real transfer, the actuator layer itself is learned. A transformer encoder with 2 layers, 4 attention heads, and history length $1.08^\circ$6 at 20 Hz maps temporal motor histories $1.08^\circ$7 to tendon force estimates, which are then embedded into a GPU-accelerated rigid-body simulator for PPO training (Yuryev et al., 4 Mar 2026).

Cosserat-rod methods provide a more mechanics-complete description for RTRs whose tendon paths, stiffness distributions, or lumen constraints matter explicitly. The catheter model represents both catheter and tendon as Cosserat rods coupled by penalty-based lumen and endpoint constraints, and uses a stable implicit Euler method to handle stiff elastic and constraint forces (Villard et al., 2024). The tapered TPU continuum robot extends tendon-actuated Cosserat formulations to spatially varying cross-section, with stiffness matrices $1.08^\circ$8 and $1.08^\circ$9 that depend on the local area and area moments of inertia and their derivatives, thereby capturing the graded stiffness induced by taper (Hansen et al., 19 Mar 2026).

5. Experimental demonstrations and quantitative performance

The cited literature reports that reconfiguration materially changes performance rather than only broadening design vocabulary. In the bead-based variable-stiffness manipulator, internal routing reduced mean proximal deflection during distal actuation from 6.55 mm, 10.34 mm, and 15.86 mm under external routing to 0.53 mm, 1.39 mm, and 5.52 mm, corresponding to relative reductions of 93.74%, 89.49%, and 70.11% across distal curvatures of $S$0, $S$1, and $S$2. Internal routing combined with stiffening reduced those values further to 0.09 mm, 0.09 mm, and 0.23 mm. In the out-of-plane load case $S$3, the manipulator supported 1 kg with only about 13 mm tip deviation, while in the horizontal configuration both in-plane and out-of-plane cases reached a threshold of about 200 g and the maximum reach decreased by only 20% relative to the vertical configuration (Walker et al., 2024).

The compliant snake robot shows how routing-mode reconfiguration changes locomotion rather than static shape alone. On a tiled indoor floor, the measured speeds were 27.6 mm/s forward, 35.5 mm/s backward, and 20.0 mm/s in sidewinding, and the robot was also shown turning a $S$4 corridor corner. The paper is explicit that the dominant failure mode was flipping over, caused by torsional backlash and insufficient torsional stiffness of the hand-built continuum joints (Kwon et al., 2024).

The lockable-joint RTR reports performance gains tied to coupling elimination and workspace restructuring. Its workspace analysis proves $S$5 for the compared idealized systems, and its dexterity study reports maximum planar dexterity values of 21.85% for 3 divisions, 55.72% for 4 divisions, 61.50% for 5 divisions, and 66.75% for 6 divisions while keeping total length constant. Static-model validation on the seven-joint prototype, with friction coefficient calibrated to $S$6, yielded worst mean posture error and standard deviation across joints below $S$7 (Lin et al., 23 Jul 2025).

Learning-based RTR control also yields quantified benefits. For SoftQ, the trained MBRL policy converged in about 200 episodes using the surrogate model, MBRL+PT converged in roughly 60 episodes in the physical simulator, and the model-free baseline required about 450 episodes while the prior benchmark needed around 600 episodes or more. MBRL+PT achieved about $S$8 in simulation, corresponding to 1.8 m in 5 s, and about 0.13–0.15 m/s in real hardware; the benchmark table reports stability score 0.65 for MBRL+PT versus 0.03 for MFRL, COT 68 J/kg/m versus 501 J/kg/m, average walking speed 0.26 m/s versus 0.003 m/s, and training time 10.8 h versus 22.6 h (Niu et al., 2024). For tendon-force-aware sim-to-real transfer, the transformer-based force model achieved average RMSE 0.61 N across weak spring, strong spring, and finger setups, corresponding to 2.9% of the motor’s maximum force of 21 N; under sinusoidal fingertip tracking, ideal-force simulation had 14.58 mm RMSE while the transformer-based simulation had 8.61 mm RMSE, a reported 41% improvement, and real-robot RL tracking improved from 24 mm RMSE to 12 mm RMSE, a reported 50% improvement (Yuryev et al., 4 Mar 2026).

Medical and soft-continuum RTR variants show similar patterns. In the separable catheter TDRM, hysteresis compensation reduced straight-condition bending-angle error from $S$9 to $S$0, a reported 46.6% reduction; under curved proximal-section tests, combined re-tension and hysteresis compensation reduced average xy-plane error over $S$1, $S$2, and $S$3 trials to $S$4, a reported 52.35% reduction relative to baseline; and in the dynamic proximal-angle-change condition, final configuration error was reduced by 89.14% (DeBuys et al., 2023). In the 2019 layer-jamming continuum robot, the stiffness ratio of the two-segment arm increased from 1 to 17.5 as vacuum pressure increased from 0 to 12.5 psi, and the single-segment stiffness ratio increased from 1 to 9.2 (Ouyang, 2019).

6. Applications, limitations, and open technical questions

The application domains represented in this corpus are diverse but mechanically coherent. The bead-based variable-stiffness manipulator targets larger-scale continuum designs and environments with very different hydrostatic pressures, where pneumatic chambers are problematic (Walker et al., 2024). The snake robot uses reconfiguration of whole-body contact patterns for forward motion, backward motion, and sidewinding (Kwon et al., 2024). The tapered TPU platform supports teleoperated grasping and endoscopic gripper integration on a 6-DoF arm (Hansen et al., 19 Mar 2026). The catheter literature targets minimally invasive cardiac procedures, where accurate shape and force prediction are needed for control and for interaction with tissue and vasculature (Villard et al., 2024). The origami-inspired manipulator and the lockable-joint RTR focus on obstacle avoidance, target alignment, and workspace-efficient maneuvering in cluttered environments (Chen et al., 14 Apr 2025, Lin et al., 23 Jul 2025).

Several recurring technical tensions define the RTR research agenda. The first is compliance versus stability. The cited work rejects the assumption that soft or continuum bodies must sacrifice payload and tip stability: tendon-induced stiffening, internal routing, and locking strategies are all presented as ways for flexibility and tip stability under loading to co-exist without compromise (Walker et al., 2024, Lin et al., 23 Jul 2025). The second is reconfigurability versus controllability. Active rerouting, high-dimensional actuation spaces, torsional backlash, tendon friction, hysteresis, and sim-to-real actuator mismatch all make inverse control harder, which is why the literature increasingly relies on curvature–torsion signatures, morphology-aware action spaces, learned force models, and co-optimization of design and control (Dash et al., 14 Apr 2026, Kwon et al., 2024, Yuryev et al., 4 Mar 2026, Niu et al., 2024, Chen et al., 14 Apr 2025).

A second misconception is that “model-free” RTR control means an absence of geometric or mechanical structure. The RTDCM shape-matching framework is described as a model-free alternative because it bypasses a full explicit mechanical inversion, yet it still relies on experimentally observed curvature–torsion signatures and sequential one-dimensional optimization (Dash et al., 14 Apr 2026). Likewise, reinforcement learning approaches do not operate on arbitrary motor spaces when they are most effective; they operate on reduced morphology-aware spaces or on simulations enriched with learned tendon-force models (Niu et al., 2024, Yuryev et al., 4 Mar 2026).

The literature is also explicit about limitations. The snake robot reports modest speed and flipping due to insufficient torsional stiffness (Kwon et al., 2024). The origami co-optimization results are in simulation and assume continuously adjustable stiffness, although earlier work only demonstrated binary stiff/soft states (Chen et al., 14 Apr 2025). The catheter Cosserat model uses penalty constraints rather than exact constraints and does not yet model friction against lumen walls or contact with vasculature and tissue in the validated form (Villard et al., 2024). The 2019 layer-jamming prototype fabricated only two segments instead of the intended three, and the latex vacuum bag reportedly degraded after about 100 cycles (Ouyang, 2019). The separable medical manipulator characterized proximal-section effects mainly for one-dimensional bending, leaving multi-bend shapes such as S-curves for future work (DeBuys et al., 2023).

Taken together, these works define RTR as a technically specific class of tendon-driven robots whose distinctive property is a reconfigurable actuation-to-shape relationship. That relationship may be changed by modular length, tendon routing, stiffness, joint locking state, proximal configuration, or learned control abstraction. The corpus suggests that the central challenge is not only how to deform a compliant body with tendons, but how to redesign or reinterpret the tendon-body system so that the same hardware can assume different mechanically meaningful roles under different tasks and environments.