Modularized Artificial Muscular System (MAMS)

Updated 15 November 2025

Modularized Artificial Muscular System (MAMS) is a reconfigurable robotic framework that integrates standardized muscle and skeletal units to mimic biological musculoskeletal architectures.
It employs diverse actuation methods, including LCE rods, tendon-driven motors, and HASEL actuators, to achieve high adaptability, distributed control, and rapid morphological reconfigurability.
Advanced computational design workflows and adaptive control strategies enable precise real-time performance, plug-and-play upgrades, and scalable robotic applications.

A Modularized Artificial Muscular System (MAMS) is a robotic hardware and computational framework for constructing, actuating, and controlling reconfigurable musculoskeletal structures using modular artificial muscle and skeletal units. These systems are designed for high adaptability, distributed actuation, and physical configurability, drawing direct inspiration from biological musculoskeletal architectures. Current research on MAMS encompasses hard robotic limbs with tendon-driven actuation, soft structures emulating biological hydrostats, and hybrid platforms combining multiple artificial muscle technologies. The following sections provide a technical overview of the constituent modules, actuation models, system architectures, control methodologies, computational design workflows, and experimental benchmarks of state-of-the-art MAMS implementations.

1. Modular Hardware Architectures

MAMS hardware consists of a composition of functionally distinct, physically standardized modules enabling flexibly reconfigurable robotic bodies:

Skeletal Modules: 3D-printed bones (e.g., TPU-95A for flexibility or PLA for rigidity) or CNC-milled aluminum “bone frames” define the mechanical scaffold. Geometry includes end-caps for rotational joints and slot features to accommodate muscle-induced shrinkage or tendon routing (Bhargava et al., 7 Aug 2025, Kawaharazuka et al., 29 Oct 2024, Kawaharazuka et al., 10 Nov 2024, Yuan et al., 8 Nov 2025).
Artificial Muscle Modules:
- Contractile Rods: Liquid crystal elastomer (LCE) rods capable of up to 33% thermally induced contraction (1 N force per rod under 808 nm laser, cycling between 25°C–120°C), glued to bones for direct actuation (Bhargava et al., 7 Aug 2025).
- Tendon-Driven Actuators: Brushless DC motor modules winding Dyneema or polyester fiber tendons, with integrated tension sensing and mounting flanges for rapid reattachment (Kawaharazuka et al., 29 Oct 2024, Kawaharazuka et al., 10 Nov 2024, Yuan et al., 8 Nov 2025).
- Soft/Electric Muscles: Modular series of HASEL electrohydraulic actuators (open-loop bandwidth >3 Hz) integrated with electrostatic clutches for antagonistic joint action (Kazemipour et al., 17 Sep 2024).
Joint and Routing Modules: Standardized joint modules (1–3 DoF via interchangeable axes with onboard sensing) and routing units (relay pulleys, nonlinear elastic units) facilitate arbitrary network topologies (Kawaharazuka et al., 29 Oct 2024, Kawaharazuka et al., 10 Nov 2024).
Mechanical/Electrical Interface Standards: Universal through-hole matrices (10 mm–20 mm grid), keyed attachments, and plug-and-play electrical connectors enable reassembly, hot-swap, and expansion without system redesign (Kawaharazuka et al., 29 Oct 2024, Kawaharazuka et al., 10 Nov 2024, Yuan et al., 8 Nov 2025).

This modular approach supports rapid reconfiguration into new morphologies or body plans (full humanoids, single limbs, soft tentacles), scalability from small-scale testbeds to large humanoids, and straightforward incremental hardware upgrades.

2. Actuation Physics and Module Modeling

The actuation behavior of MAMS modules is governed by the specific artificial muscle implementation and mechanical transmission design:

Liquid Crystal Elastomer (LCE) Actuators (Bhargava et al., 7 Aug 2025):
- Contraction ratio $r(T) = (L_0 - L(T))/L_0$ follows a sigmoidal dependence on temperature, with up to $r_{max} \approx 0.33$ .
- Lumped-parameter heat balance: $\rho c A_m \frac{dT}{dt} = \alpha I(t) - h(T-T_0)$ .
- Muscle force-strain: $F(\varepsilon) = k_m \cdot \varepsilon$ , $k_m \approx 30$ N.
Tendon-Driven Motor Muscles (Kawaharazuka et al., 29 Oct 2024, Yuan et al., 8 Nov 2025):
- Hill-type muscle model: $F^m = a\,F_0^m\,f_\ell(\bar\ell^m)\,f_v(\dot{\bar\ell}^m) + f_{pe}(\bar\ell^m)$ , with parallel and series elasticity modeling muscle and tendon elements.
- Activation dynamics: $\dot a = \frac{u - a}{t_a(u,a)}$ , reflecting different rise/fall time constants for actuation/deactuation.
Electrohydraulic and Clutch Modules (Kazemipour et al., 17 Sep 2024):
- HASEL actuator static force: $F_{actuator}(\Delta x;V) \approx F_0(V)(1 - \frac{\Delta x}{\Delta x_{max}(V)})$ .
- Electrostatic clutch holding stress: $P_c(V)$ , with engagement torque $T_{clutch}(V) = P_c(V) \cdot A_{overlap} \cdot r$ .
Soft Continuum and Hydrostat Models (Tekinalp et al., 2023):
- Continuum mechanics (Cosserat rods) for muscular hydrostats; individual rod modules parameterized for stretch, bend, and twist with task-relevant topological invariants (link, writhe, twist).
- Modular activation “templates” specify spatiotemporal contraction profiles per muscle group.

Associated actuation models provide closed-form expressions for deformation and output force, supporting both predictive simulation and real-time control law synthesis.

3. System-Level Assembly, Sensing, and Control

Mechanical Assembly: Modular bone, joint, and routing units combine into arbitrary networked “skeletons” specified as graphs $G=(V,E)$ , supporting 1D chains, 2D meshes, or complex 3D bodies (Bhargava et al., 7 Aug 2025, Kawaharazuka et al., 29 Oct 2024).
Sensing: Joint and muscle modules embed rotary encoders, IMUs, tension/load cells, and—in LCE systems—external ArUco markers for optical 6D tracking and feedback (Bhargava et al., 7 Aug 2025, Kawaharazuka et al., 29 Oct 2024, Kawaharazuka et al., 10 Nov 2024).
Electrical and Network Topology: Modules expose device endpoints over USB, CAN, EtherCAT, or HV bus; device discovery and actuation commands managed by host PCs with automated enumeration (Kawaharazuka et al., 29 Oct 2024, Kawaharazuka et al., 10 Nov 2024, Yuan et al., 8 Nov 2025).
Control Loops:
- Local motor current/tension or position control at 1–2 kHz for tendon modules.
- Recursive estimation and adaptive update of whole-body self-image (body schema) using joint and muscle sensor fusion in real time (Kawaharazuka et al., 29 Oct 2024, Kawaharazuka et al., 10 Nov 2024).
- For soft/optically-stimulated systems, untethered distributed actuation is implemented by steering collimated IR lasers with galvoscanners and tracking contraction via vision (Bhargava et al., 7 Aug 2025).
- Learning-based adaptive control: online model learning (incremental least squares; self-image update), dual-term loss to preserve previous knowledge after muscle addition, or iterative learning (DDILC) for repetitive tasks (Kawaharazuka et al., 10 Nov 2024, Yuan et al., 8 Nov 2025).
- Hybrid state machine controllers for actuator–clutch pairs (e.g., HASEL-clutch) to coordinate contraction and extension phases per antagonistic joint (Kazemipour et al., 17 Sep 2024).

A key architectural feature is system-wide plug-and-play reconfigurability, mechanical and electrical standardization, and the ability to grow or shrink the set of actuators with minimal software changes.

4. Computational Design and Optimization Workflows

Two principal computational methodologies optimize MAMS for desired performance targets:

Skeletal Graph and Shape Morphing Optimization (Bhargava et al., 7 Aug 2025):
- Given $N$ target shapes $\{\Omega_0^i\}$ , optimize a 2D “blueprint” graph $\Gamma$ and 3D configurations $\Omega^i$ subject to mechanical, fabrication, and actuation constraints.
- Mixed discrete-continuous energy minimization with objectives for bijectivity, conformal distortion, mesh quality, and approximation accuracy. Soft constraints ensure feasible contraction ranges and total actuation budget.
- Solution proceeds via iterative discrete (edge flips/splits/collapses) and continuous (vertex position) updates.
Joint Skeleton–Control Gait Co-optimization:
- Simultaneous search over skeleton vertex positions $V$ and per-edge periodic muscle activation parameters (amplitude, frequency, phase) for locomotion objectives (speed, incline, fluid motion).
- Pipeline: global coarse search (differential evolution) followed by alternating refinement and simulation-based validation (MuJoCo).
Adaptive Body Schema Learning and Incremental Muscle Addition (Kawaharazuka et al., 10 Nov 2024):
- Expand the existing joint/muscle state–mapping network to accommodate added actuators; retrain with a small number of new data points using a dual-term loss, copying weights from previous network for continuity.
Iterative Learning Control (ILC) for Dynamic Repetitive Tasks (Yuan et al., 8 Nov 2025):
- MIMO time–iteration learning laws update both feedback and feedforward components to minimize trajectory error across repeat trials.

These computational tools greatly streamline both physical assembly (by hardware-aware CAD export from skeletal graphs) and controller synthesis, supporting task-specific and general-purpose robot morphologies.

5. Experimental Performance and Benchmarking

Empirical studies across multiple MAMS platforms validate both the modular assemblies and their adaptive performance:

Platform	Actuator Type	Task/Metric	Key Results
LCE–Bone system	LCE rods/laser	Morphing & locomotion	±90° bending, 33% contraction in 1–2 s, flat walk at 1.2 cm/s, incline climbing, water swim
Musashi arm	DC/motor+tendon	Force absorption, learning	60 N impulse absorption, 3.6 kg lift (25 Nm joint), RMS < 2° with online learning
HASEL+Clutch	Electrohydraulic/ES	Antagonistic joint, range of motion	±82° ROM at 2.5 Hz, 1.5 Nm torque, 15 ms cycle, rapid module exchange
Tendon-driven arm	DC/motor+tendon	7DOF, ILC, tracking under disturbance	1.4 mm (0.14%) 3D error after 90 iters, <1% error for ≤20% load disturbance, robust to 15 actuators
Sarcomere-mimic	Pneumatic soft actuator	Force/displacement modular scaling	0.315–1.03 N (1–3 unit), ~35% strain, 18% force synergetic gain for 3-APS myofibril

Further, multi-muscle addition under adaptive learning sees per-muscle tension decrease by 45% in high-load tasks, and modular skeletons are reconfigured for new limb types or morphing gaits without software changes. This supports claims of distributed adaptability, robustness, and plug-and-play scalability.

6. Prospects, Limitations, and Design Guidelines

Current MAMS research demonstrates significant strengths in adaptability, rapid reconfigurability, and learning-based task adaptation:

Scalability: Modular physical and electrical interfaces (uniform hole patterns, bus/topology networks) support robot expansion to large DOF and node counts (Bhargava et al., 7 Aug 2025, Kawaharazuka et al., 10 Nov 2024, Yuan et al., 8 Nov 2025).
Plug-and-Play Upgrades: Muscle modules are hot-swappable (~3 min/module for Musashi), and extension to heavier DOF supported by parallel motorization or module scaling (Kawaharazuka et al., 29 Oct 2024).
Distributed, Untethered Control: LCE systems avoid accumulative wiring, with distributed activation via optical tracking and laser steering (Bhargava et al., 7 Aug 2025).
Rapid Computational Redesign: Highly automated pipelines enable CAD to print to assembly workflows, with only manual glue/marker placement.
Soft–Rigid Integration: Both rigid and soft/continuum architectures are achievable under unified modular abstractions (e.g., Cosserat module bundles for soft arms (Tekinalp et al., 2023)).
Limitations:
- Heating/cooling and contraction power budgets restrict current LCE-actuated assemblies (≈30 muscles at full range) (Bhargava et al., 7 Aug 2025).
- Tuning and calibration of soft actuators require feedback for manufacturing variability (Bhargava et al., 7 Aug 2025, Labazanova et al., 2021).
- Control bandwidth of soft actuators (<5 Hz) may limit high-frequency tasks (Kazemipour et al., 17 Sep 2024).
- Cumulating geometric/fabrication errors in series/parallel soft muscle chains (Labazanova et al., 2021).

Design Guidelines:

Standardize mechanical footprints and connectors for all modules.
Select muscle modules according to estimated joint load and performance profile; for high-load, double-up or select higher-gear motor modules.
Tune nonlinear or compliant elements to application (Grommet-NEU for high load, Oring-NEU for low load) (Kawaharazuka et al., 29 Oct 2024).
Maintain a modular body-schema network capable of dimension expansion and rapid post-upgrade adaptation (Kawaharazuka et al., 10 Nov 2024).
For large humanoids, scale modules (e.g., φ28–φ35 mm BLDC, higher-torque gearboxes) (Kawaharazuka et al., 29 Oct 2024).
For soft systems, manage cumulative error by precision casting/fabrication and run early feedback-based calibration (Labazanova et al., 2021, Bhargava et al., 7 Aug 2025).

This suggests that future advances in MAMS will likely combine distributed soft/rigid actuation, self-adaptive schema learning, and modularity at both mechanical and computational levels, supporting highly adaptable and robust robotic agents spanning the spectrum from continuum manipulators to articulate humanoids.