Modular Artificial Muscular System (MAMS)

• Modular Artificial Muscular Systems are reconfigurable actuator–sensor modules that emulate biological muscles and tendons to achieve high dexterity and adaptability.
• MAMS integrates standardized modules, including electric, pneumatic, and continuum actuators, enabling rapid plug-and-play assembly, robust load sharing, and dynamic compliance.
• Experimental results indicate sub-percent tracking errors, a 45% reduction in peak tension through added modules, and reconfiguration times under 30 minutes, demonstrating practical scalability.

A Modular Artificial Muscular System (MAMS) is an architectural, mechanical, and control paradigm for constructing artificial musculoskeletal structures from discrete, reconfigurable, actuator–sensor modules, inspired by biological muscle, tendon, and hydrostatic systems. MAMS architectures aim to achieve compliant, dexterous, high-DOF actuation for advanced robotics, prosthetics, and bioinspired mechanisms, while facilitating incremental hardware/software adaptation, scalability, and maintenance through modularity. Research on MAMS encompasses electric tendon-driven systems, pneumatic networks, continuum muscle arrays, and learning-enabled body schemas, spanning humanoid robots, octopus-inspired hydrostats, and printed musculoskeletal hands (Yuan et al., 8 Nov 2025, Kawaharazuka et al., 2024, Tekinalp et al., 2023, Buchner et al., 2024, Labazanova et al., 2021, Kawaharazuka et al., 2024).

1. Modular Architectures and Hardware Topologies

MAMS designs typically distinguish themselves through standardized, hot-swappable modules for both actuation and structural connection. Muscle modules (tendon-driven, pneumatic, soft continuum, or electric) are physically and electrically interchangeable, permitting rapid assembly, replacement, and evolution of morphology.

• Skeletal Frame & Joint Modules: Platforms such as Musashi (Kawaharazuka et al., 2024), LTDM-Arm (Yuan et al., 8 Nov 2025), and 3D-printed hands (Buchner et al., 2024) deploy generic aluminum or printed bone frames, assembled with modular joint units. Joint modules provide integrated angle sensing (potentiometers, IMUs) and accept multiple muscle attachment points. Standardized geometric grids and quick-disconnect brackets support reconfiguration of degrees of freedom and muscle routing with minimal disassembly time (<30 min for limb rebuild in Musashi).
• Actuator Modules: Muscle modules vary by application:
  • Electric Tendon-Driven: Sensor-driven motor–cable units (e.g., 15-module, 7-DOF LTDM-Arm (Yuan et al., 8 Nov 2025); up to 500 N tension, 20–100 Hz bandwidth (Kawaharazuka et al., 2024)).
  • Pneumatic Artificial Muscles: Modular McKibben-style actuators (30.1% strain, 38 N blocking force (Buchner et al., 2024)) or bioinspired sarcomere-mimetic units in series (modular artificial myofibrils, 0.3–1.0 N per APS (Labazanova et al., 2021)).
  • Soft/Continuum:
  • Cosserat-rod assemblies model muscle arrays in octopus-arm MAMS analogs, discretized over ∼200 contractile elements (Tekinalp et al., 2023).
• Routing Relays and Elastic Units: Modular relay pulleys and nonlinear elastic elements provide arbitrary 3D tendon pathways and tunable compliance. NEUs (nonlinear elastic units) confer variable stiffness for antagonistic control and facilitate compliance tuning at the module level (Kawaharazuka et al., 2024).
• Sensing & Communication: Tension sensors (load cells), position encoders, integrated microcontrollers, and networked data buses (USB, CAN, RS-485) are modular and daisy-chainable. Actuator state and sensor data are uniformly available to high-level controllers.

2. Actuation Technologies and Biomechanical Models

MAMS includes a spectrum of actuator modalities, each paired with a principled biomechanical or continuum-mechanical model for force, compliance, and control synthesis.

• Hill-Type Muscle Models: Adopted for both cable-driven (Yuan et al., 8 Nov 2025) and musculoskeletal humanoid platforms (Kawaharazuka et al., 2024). A canonical form is:

$$F^m = a\,F^m_0\,f_\ell(\bar{\ell}^m)\,f_v(\bar{v}^m) + f_{pe}(\bar{\ell}^m)$$

with experimentally parameterized force–length and force–velocity relations.

• McKibben/Pneumatic Actuation: PAMs characterized by approximate linear force–strain relations (e.g.,

$F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$

with $k$ as fit coefficient), block force $\sim$38 N at 0.5 MPa (Buchner et al., 2024). Series arrangements of APS modules allow direct force/length scaling (Labazanova et al., 2021).

• Continuum Rod Models: For hydrostatic MAMS, the assembly is modeled as a bundle of Cosserat rods with distributed internal force/moment balances. Constitutive and active muscle terms drive overall kinematics:

$$\partial_s n + f_{ext} = \rho A \partial^2_t r \ \partial_s m + r' \times n + l_{ext} = \rho I \partial_t \omega$$

Activation templates $a_g(s,t)$ define the internal torque distributions for motion primitives (Tekinalp et al., 2023).

• Nonlinear Elastic Units: NEU response is exponential in tension–elongation, e.g. $T = a_m \exp(b_m \Delta l_m)$, enabling tunable module-level compliance (Kawaharazuka et al., 2024).

3. Modularity: Integration, Scalability, and Reconfiguration

Key to MAMS is a physically and functionally modular framework, enabling:

• Plug-and-Play Assembly: All elements—actuators, joints, bone frames, relays, elastic elements—are joined via standard mechanical and electrical interfaces, supporting local expansion or substitution. Whole-limb rebuilds or DoF changes are achievable in under 0.5 h (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024).
• Incremental Morphological Growth: Muscle modules can be added wherever increased DoF, torque capacity, or compliance is needed. Skeletal extension, relaying, and muscle redundancy are achieved by module stacking, paralleling, or flexible tendon/plumbing rerouting (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024).
• Self-Calibration and Fault Tolerance: Each module's integrated sensing supports in situ calibration and self-identification for adaptable control allocation. Redundant muscle paths or antagonistic pairs maintain operation under unit failure (Yuan et al., 8 Nov 2025).

Table 1: Core MAMS Module Types and Representative Parameters

Module Type	Force/Torque Capacity	Typical Sensing/Control
DC tendon-motor act.	up to 500 N, 2.5 Nm	Load cell, position encoder
Pneumatic artificial muscle	1–38 N blocking	Pressure sensor, valve servo
Cosserat continuum segment	<5 N per element	Flex/twist sensor, strain gage
Nonlinear elastic unit (NEU)	up to 300 N	Strain measurement, passive

4. Control Architectures and Adaptive Body Schema

MAMS systems leverage modular control hierarchies which map high-level kinematic goals to distributed module activations, with adaptive, learning-enabled body schemas for rapid topology changes.

• Hierarchical Controllers:
  • High-level planners specify trajectories in joint, end-effector, or topological coordinates (e.g., $\Delta Lk, \Delta Wr, \Delta Tw$ in octopus arm (Tekinalp et al., 2023)).
  • Mid-level mappers compute muscle activation templates via traveling waves, pulses, or uniform ramps for dynamic behavior composition.
  • Low-level loops perform tension/pressure servo or direct activation tracking at each module (typically 20–1000 Hz) (Yuan et al., 8 Nov 2025, Kawaharazuka et al., 2024).
• Learning-Based Body Schema:
  • Musculoskeletal AutoEncoder (MAE): A neural network simultaneously modeling all mappings among joint angles, muscle forces, and lengths, allowing arbitrary addition of new muscle modules with minimal retraining via weight-copy and incremental data (Kawaharazuka et al., 2024).
  • Online Parameter Identification: Recursive least-squares/gradient schemes adapt geometric and elastic model parameters during normal operation, compensating for friction, compliance, or assembly tolerance drifts (Kawaharazuka et al., 2024).
  • Data-Driven Iterative Learning Control (DDILC): Repetitive tasks refine MAMS activation profiles via feedforward update laws,

  $u_{k+1}(t) = u_k(t) + L\,e_k(t)$

  achieving sub-millimeter accuracy, sub-percent repeatability under load (Yuan et al., 8 Nov 2025).

• Rapid Body Schema Expansion: Adding muscle modules triggers weight copying in MAE and minor new data collection, achieving tension relaxation (peak reductions up to 45%), more even load sharing, and preservation of kinematic accuracy after just 10–30 small-data retraining epochs (Kawaharazuka et al., 2024).

5. Experimental Performance and Evaluation

MAMS-based robots have been evaluated across human-scale arms, printed hands, continuum appendages, and task-specific manipulation:

• Trajectory Tracking Robustness: The LTDM-Arm maintains $<1.5\%$ error under up to $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$0 load disturbances (simulation: $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$1 error for up to $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$2 load), with DDILC outperforming CMC, MTIS-FCC, and nonlinear-PID by 58-78% in mean square error (Yuan et al., 8 Nov 2025).
• Load Sharing: Adding new muscles allowed a 45% reduction in single-muscle peak tension in high-load arm lifts on the Musashi platform, while retaining joint angle error within a few degrees (Kawaharazuka et al., 2024).
• Compliant Dexterity: 22-PAM printed hands produced 30% contraction strain and 38 N blocking force, achieving successful independent finger and power/precision grasps (measured fingertip force: $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$3 N, grasp: $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$4 N) (Buchner et al., 2024).
• Continuum Control: Octopus-arm analogs demonstrated high-DOF manipulation through superposition of muscle activation templates; topological invariants $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$5 coordinate wrap/grasp routines (Tekinalp et al., 2023).
• Adaptive Learning: Handle-rotation tasks on Musashi reduced tension tracking error from 50 N to $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$6 N within $F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$7100 learning steps, outperforming classic PID by range and load at lower effort (Kawaharazuka et al., 2024).

6. Theoretical and Design Formulations

Foundational MAMS equations derive from continuum rod theory, coupled actuator models, and learning-driven body representation.

• Cosserat Rod Framework for Continuum MAMS (Tekinalp et al., 2023):

$F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$8

with kinematic (5a-c), topological, and activation template equations.

• Muscle Actuator Modules (Yuan et al., 8 Nov 2025, Kawaharazuka et al., 2024):

$F(P, \epsilon) = k P (1-\epsilon/\epsilon_{max})$9

with explicit modular empirical/analytical fits for PAMs (Buchner et al., 2024, Labazanova et al., 2021).

• Control Templates and Adaptive Schema (Kawaharazuka et al., 2024):

$k$0

7. Implications, Limitations, and Future Prospects

MAMS enables new classes of robotic embodiments—modular, resilient, and amenable to autonomous adaptation. Modular architectures offer:

• Hardware Scalability and Adaptation: Rapid, local changes in actuation topology, tailored stiffness, and task-specific redundancy are achievable without reengineering entire systems.
• Incremental Software Generalization: Unified neural and analytical body schemas handle actuator growth, sensor failures, and dynamic reconfiguration based on minimal new data.
• Distributed Sensing and Compliance: Intrinsically compliant modules enable robust performance in uncertain environments; nonlinear elastic elements and muscle co-contraction support variable impedance and shock absorption.

Limitations include dependence on mechanical interface engineering for seamless scaling (especially plumbing in pneumatic implementations), increased physical footprint for high-force modules, and batch-variability in soft material properties for pneumatic and continuum units. Trade-offs between force output, displacement, and fabrication complexity are prominent in sarcomere-mimetic and PAM series. Complexity grows with the number of integrated modules in both control and sensing.

A plausible implication is that future MAMS frameworks will converge on unified modular standards—mechanical, electrical, and network—facilitating large-scale adaptive robotic organisms, sophisticated body-schema learning, and closed-loop interaction with unstructured environments, with potential translation to neuroprosthetics, soft exosuits, and human–robot symbiosis.