Tendon-Driven Anthropomorphic Manipulators

• Tendon-driven anthropomorphic manipulators are robotic systems that mimic human tendon-actuation by routing flexible cables through anatomically inspired joints for precise and compliant motion.
• They employ diverse actuation methods—such as electric motors, twisted string actuators, and pneumatic muscles—to achieve high dexterity with underactuation and antagonistic configurations.
• Advanced kinematic modeling, state estimation, and closed-loop control enable these systems to perform adaptive, robust manipulation in complex real-world tasks.

Tendon-driven anthropomorphic manipulators are robotic hands and arms whose actuation relies on flexible tendon-like elements—typically cables or strings—that are routed through anatomically inspired joints to transmit force from actuators to end-effectors. This principle, abstracted from musculoskeletal human anatomy, enables high joint count with remote actuation, reduced moving mass, compliance, and dense packaging. Modern designs exploit advanced tendon routing, underactuation, bioinspired transmission, and elastic elements to realize dexterous, adaptive, and robust human-like manipulation for both research and practical applications.

1. Architectural Principles and Kinematic Foundations

Tendon-driven anthropomorphic manipulators implement force transmission and joint actuation via cable, string, or flexible wire elements routed through routing guides, pulleys, or anatomical passages mimicking biological tendon paths. Actuators—electric motors, twisted-string actuators (TSAs), pneumatic muscles, or linear servos—are often located proximally (in the palm, forearm, or base) and remotely drive distal phalanges, wrists, or elbows by modulating tendon tension or length. Canonical kinematic schemes for these manipulators include fully actuated, underactuated, and antagonistic pair layouts:

• Fully actuated parallel transmission: Each DOF has an independent actuator; cables drive the flexion and extension of each joint. The IRMV hand demonstrates a parallel tendon-driven MCP and PTFE-conduit routing for DIP/PIP (Han et al., 4 Dec 2025).
• Underactuated synchronous routing: Multiple joints or fingers move with fixed angular ratios under a single actuator. The UTRF finger enforces θᵢ = (R₁/Rᵢ)·θ, imposing fixed coupling ratios (Yuan et al., 11 Dec 2025). The Tactile SoftHand-A and Educational SoftHand-A use differential and clutch elements for underactuated synergy (Li et al., 2024, Lepora et al., 17 Oct 2025).
• Antagonistic arrangements: Agonist/antagonist tendons allow for joint stiffness modulation and compliance adjustment, as in TSA antagonistic pairs (Bombara et al., 2022), series elastic arrangements (Lee et al., 16 Sep 2025), and classical musculoskeletal configurations (Makino et al., 2024).

Kinematic mapping from joint angles $\theta$ to tendon lengths $l$ is generally nonlinear, especially with complex routing and branching. Commonly, the tendon–joint relationship is modeled as $l = f(\theta)$, with the coupling (moment-arm) matrix $C(\mathbf q) = \partial f/\partial q$ defining how tendon excursions produce joint motion (Li et al., 2024, Polcz et al., 28 Jan 2026).

2. Actuation Modalities, Tendon Routing, and Stiffness Modulation

A range of actuators and tendon architectures have been demonstrated:

• Electric motor-driven tendons: Most designs employ DC or stepper motors wind cables on spools, providing high bandwidth and precise control. Distributed layouts, with small motors in the palm for ab/adduction and flexion/extension and higher-torque ones in the forearm for PIP/DIP flexion, realize human-like dexterity within compact volume (Han et al., 4 Dec 2025).
• Twisted String Actuators (TSA): TSAs use motor rotation to twist inextensible strings, contracting them and delivering high force in compliant, compact form factors. TSA antagonistic pairs realize tunable joint stiffness, as the stiffness $K_{\alpha}$ scales with antagonist pretwist (Bombara et al., 2022).
• Pneumatic muscles: Used in arms such as PAMY2, pneumatic artificial muscles remotely actuate via Bowden-cable tendons, yielding low inertia, high force, and intrinsic compliance. Actuators are decoupled from moving links and housed at the base (Guist et al., 2023).
• Tension-amplification and base actuation: The ATDM integrates tension amplification (multiple wraps between pulleys) to increase both output force and apparent cable stiffness, while all actuators are located at the base for minimal moving inertia (Xu et al., 2024).
• Series elastic and passive spring joints: Compliance is modulated either with embedded series elastic elements (e.g., sliding-block SEAs with measured displacement) or with integrated torsion springs at each joint (machined or wire), providing both impact resilience and variable stiffness (Makino et al., 2024, Lee et al., 16 Sep 2025).

Tendon-based antagonistic schemes facilitate active, variable joint stiffness analogous to biological muscle pairs. Bending and lateral joint stiffness can be tuned by pre-tensioning antagonist tendons, yielding angular stiffness $K_\alpha$ that scales with the amount of pretwist or active preload (Bombara et al., 2022).

3. Modeling, State Estimation, and Control

Rigorous modeling is needed to address nonlinear kinematic coupling, friction, elastic deformation, and dynamic effects:

• Kinematic and dynamic models: Denavit–Hartenberg or product-of-exponentials formulations parameterize joint spatial relationships (Han et al., 4 Dec 2025, Polcz et al., 28 Jan 2026). The dynamic models integrate rigid-body inertia, tendon moment arms, viscoelastic elements, and friction, e.g.,

$$M(q)\,\ddot q + C(q,\dot q)\,\dot q + G(q) + h(q,\dot{q}) = C^T(q)\bigl(f_t - f_f(f_t, \dot{l})\bigr) + J_{ext}^T f_{ext}$$ (Li et al., 2024).

• Static and elastostatic modeling: Screw-theory and iterative statics methods reconcile cable tension, length, and joint posture under joint elasticity (including both force and length-controlled statics) (Feng et al., 15 Sep 2025). Accurate elastostatic models predict end-effector compliance, tip stiffness, and loaded deflection with sub-mm errors (Yuan et al., 11 Dec 2025).
• State estimation without joint encoders: Kinematic frameworks combined with tendon-length and tension measurement enable joint-angle estimation and feedback control without embedded encoders. Nonlinear optimization (e.g., interior-point methods) solves for $\theta$ given measured tendon displacement and tension, achieving multi-degree accuracy in high-DoF, simulated-compatible hands (Polcz et al., 28 Jan 2026).
• Control algorithms: Both model-based (e.g., sliding mode control with underlying Timoshenko beam representation in wrists) (Sulaiman et al., 11 Jan 2026) and synergy-based or impedance-modulating approaches are implemented. Differential clutch and soft synergies yield mechanical compliance and adaptation to object contacts, while explicit PID and torque-control frameworks, including proprioceptive feedback and feedforward terms, enable both stable and dynamic trajectory tracking (Han et al., 4 Dec 2025, Lee et al., 16 Sep 2025).

4. Comparative Performance, Dexterity, and Benchmarking

Anthropomorphic tendon-driven manipulators are benchmarked in dexterity, force output, compliance, load capacity, and adaptability:

• Dexterity: Comprehensive grasp taxonomy fulfillment is assessed via the Feix GRASP scheme (e.g., 31/33 or full 33/33 grasps (Bombara et al., 2022, Han et al., 4 Dec 2025)) and thumb-opposability via the Kapandji test (e.g., 6/10 points in TSA grippers).
• Load capacity and force transmission: High gripping forces are achieved (e.g., 72 N in TSA hands (Bombara et al., 2022), 80 N at the fingertip in reinforced underactuated hands (Makino et al., 2024), 11 N/finger in lightweight 15-DoF hands (Han et al., 4 Dec 2025)).
• Compliance and impact resistance: Embedded elastic elements in joints or tendon paths yield high compliance and resilience to impact, supporting robust manipulation and even self-weight support in humanoid applications (Makino et al., 2024).
• Dynamic capability: Arms such as PAMY2 achieve end-effector velocities of 2.9 m/s and maintain high position repeatability (<0.3° over 25 days), with intrinsic safety (lower collision force, up to 4× safer at given speed than traditional arms due to low inertia and compliance) (Guist et al., 2023).
• Adaptability and synergy: Highly underactuated hands (e.g., Tactile SoftHand-A) perform adaptive grasps with only two actuators, combining passive spring differentials, antagonistic tendons, and tactile feedback for secure, conformation-rich grasping across diverse objects (Li et al., 2024).

A comparative summary is given in the following table:

Manipulator Architecture	Dexterity (Feix)	Tip Force (N)	#Actuators	Special Features
TSA-driven soft hand	31/33	72	11	Tunable stiffness, silicone skin
PAMY2 pneumatic arm	N/A	N/A	6	Base actuation, safe/high-speed
IRMV 15-DoF hand	33/33	11/finger	15	Compact, human-scale, full actuation
UTRF underact. finger	N/A	3 kg tip load	1/finger	Predictable stiffness, single actuator
SoftHand-A (LEGO)	N/A	1.8/finger	2	Educational, differential clutches

5. Sensing, Proprioception, and Integration

Modern systems integrate proprioception, tactile sensing, and external feedback to close the loop in manipulation:

• Tendon-based proprioception: Series elastic actuators with direct measurement of spring deformation enable accurate estimation of tendon tension, contact timing, joint configuration, stiffness discrimination, and object classification without vision (Lee et al., 16 Sep 2025).
• Tactile integration: 3D-printed optical tactile sensors are embedded in fingertips, providing high-resolution contact and slip detection. This is leveraged in grasp stabilization and slip prevention via closed-loop state machines (Li et al., 2024).
• Joint state estimation: High-DoF, compact designs often omit joint encoders for mechanical compactness; state estimation leverages tendon displacements/tensions and efficient model inversion via nonlinear optimization (Polcz et al., 28 Jan 2026).

6. Modular, Reconfigurable, and Application-specific Manipulators

New designs address modularity, actuator count, reconfigurability, and domain constraints:

• Lockable-joint continuum designs: Reconfigurable tendon-driven continuum arms with individually lockable joints eliminate inter-segmental coupling, enabling time-multiplexed actuation, lower actuator count, and configurable workspace and dexterity (Lin et al., 23 Jul 2025).
• Distributed and centralized actuation: Manipulators designed for aerial vehicles (e.g., ATDM) use centralized actuation and tension amplification, supporting low inertia and fast, minimally disturbing operations (Xu et al., 2024).
• Application-specific adaptations: Manipulators such as Rotograb with rotating thumbs bridge biomimetic hands and industrial grippers, enlarging workspace and facilitating in-hand manipulation via kinematically decoupled tendon routing and RL-based autonomy (Bersier et al., 2024).
• Educational platforms: Open-source, LEGO-based tendon-driven hands and 3D-printed designs serve as testbeds for both pedagogy and research in tendon-driven manipulation (Lepora et al., 17 Oct 2025, Li et al., 2024).

7. Challenges, Limitations, and Future Directions

Major challenges in tendon-driven anthropomorphic manipulation remain:

• Friction and hysteresis: Tendon friction, Bowden tube nonlinearities, and stiction are prominent sources of modeling, control, and repeatability error, especially in high-tension or long-routing applications (Li et al., 2024, Guist et al., 2023).
• Complexity of routing and scalability: High-DoF hands require intricate routing and robust branch/junction modeling, particularly with tendon re-merging and underactuated schemes (Polcz et al., 28 Jan 2026).
• State estimation and sensor minimization: The drive to eliminate joint encoders for compactness necessitates further work in accurate, high-speed estimation from tendon-state sensing alone (Polcz et al., 28 Jan 2026, Lee et al., 16 Sep 2025).
• Mechanical reliability and wear: Long-term operation introduces cable stretch, creep, and wear; reliability studies show promising lifetimes but continued improvement in materials and routing design is required (Guist et al., 2023).
• Scalable control for underactuation: Underactuated and synergy-driven hands offer simplicity and adaptability but with limited independence in finger trajectories; richer hardware synergies and control polices are active areas of research (Li et al., 2024).

A plausible implication is that further advances in integrated sensing, learning-augmented control, and compliant, modular actuation will continue to push tendon-driven anthropomorphic manipulators towards higher dexterity, adaptability, and reliability for both human-robot interaction and demanding manipulation domains.