Internal Torsion Spring Compliant Actuator

• ICA is a compliant actuator that uses internal torsion springs to provide adjustable stiffness and energy-efficient actuation for modular robotic systems.
• It integrates precise torsional mechanics with sensor feedback to enable enhanced dexterity, adaptive motion control, and fault tolerance.
• The design supports rapid hardware reconfiguration and scalability, making it ideal for dynamic applications in robotics and bionics.

A Modular Artificial Muscular System (MAMS) is a reconfigurable actuation and control paradigm for robotic and bionic systems, consisting of standardized, independently operable artificial muscle units integrated with modular skeletal and sensor infrastructure. MAMS are designed to emulate the musculoskeletal architectures, compliance, and dexterous capabilities of biological systems, while supporting rapid hardware reconfiguration, incremental actuator addition, and advanced adaptive control. Implementations span cable-driven musculoskeletal robots, pneumatic muscle networks, and hydrostat-inspired morphologies, with modules ranging from microfabricated actuators with integrated sensing, to macroscopic DC-motor/wire or pneumatic artificial muscle assemblies. The modularity enables scalable, maintainable, and highly adaptable robotic structures for both research and application contexts (Yuan et al., 8 Nov 2025, Kawaharazuka et al., 2024, Kawaharazuka et al., 2024, Tekinalp et al., 2023, Buchner et al., 2024, Labazanova et al., 2021).

1. Core Architectural Principles

MAMS architectures are predicated on the principle of decomposing the actuation and sensing components of a robotic body into standardized, self-contained modules. Each module encapsulates an artificial muscle actuator (cable-driven, pneumatic, motorized, or contractile ring), often with integrated feedback elements (e.g., tension/load cell, encoder, pressure/strain sensor), and mechanical/communication interfaces that enable plug-and-play assembly onto a generic skeletal frame or hydrostat substrate (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024, Yuan et al., 8 Nov 2025, Buchner et al., 2024).

Mechanical packages in MAMS typically conform to a limited set of form factors, supporting:

• Rapid reconfiguration: Modules can be swapped, relocated, or replaced on the skeleton frame or hydrostat grid without disassembling adjacent joints or actuators.
• Scalability: Systematic duplication or insertion of modules augments degrees of freedom, force capacity, or redundancy (Kawaharazuka et al., 2024, Yuan et al., 8 Nov 2025).
• Integrated compliance: Modules often include nonlinear elastic elements (e.g., O-ring, grommet, contractile ring) or passive stretch zones, providing tunable compliance and variable stiffness at the module or joint level (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024, Tekinalp et al., 2023).
• Unified power/signal network: Modular bus architectures (USB, CAN, RS-485) enable dynamic enumeration and communication across distributed muscle/joint modules (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024).

Table – Common MAMS Module Features

Type	Actuation Mode	Sensing
Cable motor	DC/BLDC motor	Load cell
Pneumatic (PAM)	Pneumatic muscle	Pressure
SARComère unit	Pneumatic/Cable	Displacement
Cosserat rod	Contractile ring	Flex sensor
Joint module	Rotational	IMU/Encoder

2. Biomechanical and Mechanical Modeling

MAMS modules abstract the mechanical action of biological muscle–tendon units through one or more of the following formalisms:

• Hill-type muscle models: Muscle activation $a(t)$, governed by excitation $u(t)$, evolves according to

$$\dot a(t) = \frac{u(t)-a(t)}{\tau_a(u, a)},\quad F^m = F^{ce} + F^{pe}$$

where $F^{ce}$ and $F^{pe}$ are contractile and passive element forces, respectively; force–length and force–velocity dependencies are explicit (Yuan et al., 8 Nov 2025, Kawaharazuka et al., 2024).

• Cosserat rod models: Used in hydrostat-inspired systems (e.g., octopus arm analogs), each muscle bundle is discretized as a Cosserat rod parameterized by centerline $\mathbf{r}(s)\in\mathbb{R}^3$, director frame $\{\mathbf{d}_i(s)\}$, and subject to kinematic and constitutive relations

$$\mathbf{r}'(s) = \boldsymbol{\nu}(s),\quad \mathbf{d}_i'(s) = \boldsymbol{\kappa}(s)\times \mathbf{d}_i(s)\ \mathbf{n}_g = \frac{\partial U_g}{\partial \nu},\quad \mathbf{m}_g = \frac{\partial U_g}{\partial \kappa} + \mathbf{m}_g^{act}$$

with $U_g$ the elastic energy of each family (Tekinalp et al., 2023).

• Pneumatic artificial muscles (PAMs) and artificial pneumatic sarcomeres (APS): These modules are analytically modeled for force/strain generation in terms of pressure input, geometric parameters, and material nonlinearity. For McKibben-type actuators, the relation is typically

$$F(P,\varepsilon) = kP\left(1 - \frac{\varepsilon}{\varepsilon_{\mathrm{max}}}\right)$$

with $u(t)$0 a geometry/material parameter, $u(t)$1 normalized contraction, and $u(t)$2 the internal pressure (Buchner et al., 2024, Labazanova et al., 2021).

Muscle module characteristics, including peak force, displacement, bandwidth, and compliance parameters, are directly accessible through these models with empirical fitting to experimental data (Labazanova et al., 2021, Buchner et al., 2024, Yuan et al., 8 Nov 2025).

3. Modular System Integration

Assembly of a MAMS follows a systematic protocol integrating skeletal frames, joint modules, muscle modules, elastic units, and routing/relay hardware:

• Structural scaffold: Generic aluminum or 3D-printed frames with a grid of standardized attachment points enable repositioning and scaling of limb segments (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024, Buchner et al., 2024).
• Joint modules: Ball-joint, hinge, or spherical modules with integrated potentiometric or IMU-based angle sensing, providing up to three rotational DoF per module (Kawaharazuka et al., 2024).
• Muscle modules: Both “large” (high-force, sensor/driver integrated) and “small” (paired, limb-integrated) units attach directly to frames, tensioning Dyneema® cables or PAMs through designated routes (Kawaharazuka et al., 2024, Yuan et al., 8 Nov 2025).
• Elastic units: Nonlinear springs (e.g., O-ring NEU, grommet NEU) and variable compliance devices are used in series with muscles to provide bioinspired tension–elongation curves, modeled as $u(t)$3 (Kawaharazuka et al., 2024).

Standardized relay units (pulleys/folds, 4-way tension relays) are used to emulate complex tendon geometries, facilitating arbitrary 3D routing and remixing for rapid design iteration (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024).

Practical performance metrics from such assemblies include joint angle range, torque output, muscle force, cable speed, and modular bandwidth (typ. 20–100 Hz closed loop, 1 kHz network update) (Yuan et al., 8 Nov 2025, Kawaharazuka et al., 2024).

4. Adaptive and Hierarchical Control Frameworks

MAMS control leverages the modularity of actuators to permit multilevel, adaptive control strategies:

• Low-level tension/position control: Each muscle module implements an onboard current or pressure PID loop, tracking desired tension or displacement commands with direct sensor feedback (Kawaharazuka et al., 2024, Yuan et al., 8 Nov 2025).
• Inverse kinematics and pose control: Central controllers solve for muscle-length or activation targets given joint-space or end-effector objectives, using online-identified (recurrent least squares) map $u(t)$4 mapping desired joint angles or torques to muscle elongations (Kawaharazuka et al., 2024).
• Learning-based adaptive control:
  • Data-driven iterative learning control (DDILC): Feedforward muscle activation profiles are updated across repetitions to minimize tracking error, with convergence to sub-millimeter accuracy under large load disturbances (Yuan et al., 8 Nov 2025).
  • Online body-schema learning: Autoencoder-based models assimilate the relationship between joint angles, muscle tensions, and elongations, and are incrementally retrained as new actuators are added. Copy-and-retrain schemes enable scalable body adaptation while mitigating catastrophic forgetting (Kawaharazuka et al., 2024).
• Hydrostat/topological control templates: In continuum and hydrostat-inspired robots, spatiotemporally localized muscle activation templates (traveling waves, pulses, uniform ramps) enable direct manipulation of link, writhe, and twist topological quantities, facilitating grasp, alignment, and dynamic reconfiguration (Tekinalp et al., 2023).

Emphasis is placed on integrating local curvature, twist, force, and touch sensing into modules to permit closed-loop adaptation and fault-tolerant operation (Kawaharazuka et al., 2024, Buchner et al., 2024, Tekinalp et al., 2023).

5. Performance, Dexterity, and Robustness

MAMS implementations achieve performance characteristics approaching or exceeding key musculoskeletal benchmarks:

• Tracking accuracy: Sub-millimeter mean trajectory error (as low as 1.38 mm in simulation, $u(t)$5 relative error under load) with DDILC (Yuan et al., 8 Nov 2025).
• Disturbance rejection: Below 1.5% tracking error at up to 15% load in hardware, with monotonic convergence under repetitive task learning (Yuan et al., 8 Nov 2025).
• Force and grasping capabilities: PAM-based hands achieve blocking forces of $u(t)$6 N per muscle at $u(t)$7 MPa, fingertip forces of $u(t)$8 N, and grasp forces of $u(t)$9 N in modular hand designs (Buchner et al., 2024).
• Mechanical resilience: Modular sarcomere arrays (APM) retain function under single-module failure; net force scaling is approximately linear with the number of modules with modest super-additivity (Labazanova et al., 2021).
• Dexterity and adaptation: Modularity and tendon routing flexibility support anthropomorphic workspace and tasks (e.g., thumb opposition, precision grip), with learning algorithms achieving 1.6$$\dot a(t) = \frac{u(t)-a(t)}{\tau_a(u, a)},\quad F^m = F^{ce} + F^{pe}$$0 improvement in manipulation range and 33% reduction in peak muscle tension after online adaptation (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024).

6. Extension, Scalability, and Learning with Incremental Reconfiguration

A core advantage of MAMS is hot-swapping and scaling of modules to grow actuator arrays, adjust force/displacement envelopes, or tune compliance in response to new requirements:

• Plug-and-play module addition: Standard mechanical and signal interfaces permit addition or rearrangement of muscle modules in under 3 minutes per unit (Kawaharazuka et al., 2024).
• Body-schema learning for augmentation: Upon module addition, the musculoskeletal autoencoder's weights are copied to an expanded model; retraining with limited new data re-adapts the network, allowing new actuators to contribute to torque without loss of previously acquired mapping accuracy (Kawaharazuka et al., 2024).
• Load sharing and relaxation: Incremental muscle addition demonstrably reduces peak tension requirements for high-load tasks by up to 45%, distributes torque more evenly, and preserves or improves trajectory accuracy (Kawaharazuka et al., 2024).
• Global scaling: Modular assemblies can be up- or down-scaled by altering segment lengths, module power, or wiring without requiring fundamental redesign; the mechanical and control architecture generalize directly from finger to full-limb or humanoid scale (Kawaharazuka et al., 2024, Buchner et al., 2024).

7. Advanced Morphologies and Hydrostat MAMS

Expansion of the MAMS framework to continuum and muscular hydrostat morphologies broadens its scope:

• Cosserat-based hydrostat MAMS: Octopus-arm analogs discretize the arm into hundreds of modular Cosserat rod actuators, capturing 3D dynamic shape through kinematics $$\dot a(t) = \frac{u(t)-a(t)}{\tau_a(u, a)},\quad F^m = F^{ce} + F^{pe}$$1, director frame transport, and topological invariants (link, writhe, twist) (Tekinalp et al., 2023).
• Activation templates: Simple, spatially and temporally structured activation profiles (uniform, traveling-wave, pulse packet) compose into high-level behaviors—propagating bends, helical wrapping via injective twist/writhe, and robust object manipulation.
• Control law modularity: Multilevel scheme with high-level planners (targeting topological changes), mid-level activation mappers, and low-level driver circuits interfacing with networked local sensors for robust environmental interaction (Tekinalp et al., 2023).
• Sensing: Integration of fiber-optic, flex, and force sensors into each hydrostat or continuum module for full-state observation and adaptive control (Tekinalp et al., 2023).

A plausible implication is that the hydrostat modular paradigm can inform the design of soft robots requiring infinite-DoF shape control, and that continuum MAMS and discretized jointed MAMS share algorithmic principles at the level of distributed actuator coordination.

• (Yuan et al., 8 Nov 2025): Robustness study of the bio-inspired musculoskeletal arm robot based on the data-driven iterative learning algorithm
• (Kawaharazuka et al., 2024): Component Modularized Design of Musculoskeletal Humanoid Platform Musashi to Investigate Learning Control Systems
• (Kawaharazuka et al., 2024): Adaptive Body Schema Learning System Considering Additional Muscles for Musculoskeletal Humanoids
• (Tekinalp et al., 2023): Topology, dynamics, and control of an octopus-analog muscular hydrostat
• (Buchner et al., 2024): Replicating Human Anatomy with Vision Controlled Jetting -- A Pneumatic Musculoskeletal Hand and Forearm
• (Labazanova et al., 2021): Bio-Inspired Design of Artificial Striated Muscles Composed of Sarcomere-Like Contraction Units (preprint)

---

Core Equations and Summary Framework

$$\dot a(t) = \frac{u(t)-a(t)}{\tau_a(u, a)},\quad F^m = F^{ce} + F^{pe}$$2

Here, $$\dot a(t) = \frac{u(t)-a(t)}{\tau_a(u, a)},\quad F^m = F^{ce} + F^{pe}$$3 are normalized activation templates for muscle group $$\dot a(t) = \frac{u(t)-a(t)}{\tau_a(u, a)},\quad F^m = F^{ce} + F^{pe}$$4, supporting modular, compositional actuation for manipulation and dynamic body configuration (Tekinalp et al., 2023).