Robustness study of the bio-inspired musculoskeletal arm robot based on the data-driven iterative learning algorithm

AI summary

Overview

Research area: Bio-inspired robotics — musculoskeletal robot design and nonlinear control (iterative learning control for redundant, tendon-driven arms).

Technical level: Advanced. The paper assumes familiarity with robot kinematics/dynamics, Hill-type muscle biomechanics, MIMO discrete-time nonlinear systems, and dynamic-linearization-based iterative learning control.

Scope: The paper presents the design of a lightweight tendon-driven musculoskeletal arm (LTDM-Arm) and a model-free data-driven iterative learning controller, then evaluates trajectory tracking and load-disturbance rejection in simulation and on a physical prototype.

What This Paper Is About

Human arms achieve dexterity, compliance and robustness without extremely precise sensing or control, largely because of their musculoskeletal structure and redundant antagonistic muscles. Building robots that copy this structure is hard: existing musculoskeletal platforms simulate only localized joints or a few muscles, and the many redundant actuators, nonlinear muscle dynamics and rigid-flexible coupling make accurate model-based controllers difficult to derive. The goal of this work is to build a full-morphology, lightweight musculoskeletal arm and to learn its muscle activation signals directly from input/output data so that it can track trajectories and resist load disturbances without an explicit system model.

Key Contributions

1. A lightweight tendon-driven musculoskeletal arm (LTDM-Arm) with a seven degree-of-freedom (DOF) skeletal joint system (three DOF at the shoulder, one at the elbow, one at the forearm, two at the wrist), four main skeleton components (shoulder blade, humerus, ulna, radius), and a modularized artificial muscular system (MAMS) containing 15 actuators — seven muscles at the shoulder, two at the elbow, two at the forearm and four at the wrist.
2. A modularized artificial muscular system (MAMS), also described as a Motor-Cable Artificial Muscle System, combining a DC-motor-driven winch, force sensor, tendon-sheath-pulley transmission and a lubricated cable sheath (PTFE, metal tube, rubber tube) with a 0.2 mm gap between inner tube and cable.
3. A data-driven iterative learning control (DDILC) algorithm that combines along-the-time-axis biased-format feedback control with iteration-axis feedforward control, using an improved projection algorithm to estimate the partitioned Jacobian matrix, avoiding matrix inversion and requiring no system model. The paper states that convergence of the controller was demonstrated.
4. Simulation and prototype validation of robustness, including load-disturbance sweeps in simulation (0% to 30% of a rated 2.5 kg load) and disturbance experiments on the physical prototype.

Main Findings

• Simulation tracking convergence: After 60 iterations the average trajectory error fell below 2 mm (0.2%) and gradually stabilized. Once iterations exceeded 90, the average error stabilized around 1.38 mm (0.14%), with a standard deviation of approximately 0.6 mm.
• Muscle length errors: For the four joints, muscle-cluster iteration errors decreased progressively, with average and mean square errors of the overall muscle lengths of 0.053 mm and 0.062 mm respectively.
• Advantage over baseline controllers: Average trajectory errors were approximately 1.8 mm (0.18%) for DDILC, 4.1 mm (0.42%) for CMC, 6.8 mm (0.69%) for MTIS-FCC and 8 mm (0.81%) for Nonlinear-PID. Mean squared errors were 3.49 mm², 20.46 mm², 54.81 mm² and 68.62 mm² respectively. DDILC improved over the baselines by 56.1%, 73.53% and 77.5%.
• Simulated load-disturbance rejection: With disturbance loads of 0%, 5%, 10%, 15%, 20%, 25% and 30% of the rated 2.5 kg load, average trajectory errors were 3.69 mm (0.38%), 4.53 mm (0.46%), 6.01 mm (0.61%), 7.31 mm (0.74%), 8.4 mm (0.86%), 12.8 mm (1.3%) and 15.05 mm (1.53%). Disturbance beyond 25% caused significant trajectory fluctuations, but errors remained under 1% when disturbance was less than 25%.
• Prototype experiment: On the physical LTDM-Arm, the DDILC algorithm reduced error progressively; after 40 iterations the end-effector tracking error converged to 3.8 mm (0.4%). The abstract reports that the system achieves trajectory tracking under load disturbances of 20% in simulation and 15% in experiments; the truncated text of Section 5.2 does not report the numerical error values for those experimental disturbance conditions.
• Two stated mechanisms behind robustness: (1) the muscle model's internal small closed-loop — activation signal affects muscle fiber velocity, which adjusts tendon force, which feeds back through the skeleton to adjust fiber length — acts like a low-pass filter that removes pulse or high-frequency interference in the activation signal; (2) interactive force coupling among antagonistic muscles generates variable stiffness through co-contraction modulation, providing inherent compliance that passively compensates for intrinsic oscillations and external disturbances.
• Hardware benchmarking: Using comparison criteria focused on six aspects (size ratio; weight/load; skeletal similarity; muscle similarity; sensors; performance), the LTDM-Arm is reported to show advanced configuration and performance with a higher degree of human-likeness compared against the Kenshiro, Kengoro and Musashi musculoskeletal arm sections from Tokyo Institute of Technology and the Anthrob musculoskeletal arm from the Technical University of Munich.

Methodology in Plain English

The team first builds a mathematical model of the arm. Muscle behavior is represented with a Hill-type muscle model, made of a contraction element, a parallel elastic element and a series elastic element, with a pinnate angle of 1 used in the experiment. Activation follows a first-order differential equation with an activation time constant of 10 ms and a deactivation time constant of 40 ms. Force–length, force–velocity and tendon force–strain relationships are given by Gaussian and exponential functions (for example, a shape factor of 0.45 in the active force–length curve, and tendon parameters k_toe = 3, F_toe^T = 0.33, ε_toe^T = 0.609 ε_t^T, k_lin = 1.712 ε_0^T).

Joint-space redundancy is resolved with a Moore-Penrose pseudo-inverse formulation relating end-effector velocity to joint velocity and joint torque to muscle tension. The mechanical platform uses a motor-and-cable artificial muscle module in which a motor-and-reducer-driven winch adjusts cable tension, a force sensor provides real-time rope tension feedback, and a tendon-sheath-pulley system transmits force. Each DOF is driven by at least two muscles, and each muscle governs only one joint, which reduces coupling between muscle space and joint space.

For control, the nonlinear arm is written as a MIMO repetitive-motion nonlinear non-affine discrete-time system. Rather than deriving an explicit controller, the method uses tight-format dynamic linearization: the change in output is expressed as a partitioned Jacobian matrix times the change in control input, and that matrix is estimated online with an improved projection algorithm. A feedback term with a gradient-descent-updated learning gain and a feedforward term updated across iterations are added together to produce the control input, with saturation reset mechanisms applied at each step. The result is a controller that learns activation signals over repeated trials within a finite time frame using only input/output data.

Validation used a Mujoco-Python virtual environment and the LTDM-Arm prototype, with a sinusoidal trajectory in a three-dimensional workspace spanning two cycles with a period and amplitude of 200 mm and 150 mm; each trial lasted 60 seconds and was repeated 10 times.

Why This Matters

The work argues that robustness and dexterity can come partly from mechanical structure and antagonistic muscle coupling rather than only from expensive high-precision sensors and actuators, and it shows a controller that needs no analytical model of a highly redundant, nonlinear tendon-driven system — a useful direction for robots that must operate in unstructured environments.

Real-world applications:

• Humanoid and service robots operating in unstructured or unknown environments where loads and parameters are uncertain.
• Prosthetic and rehabilitation arms that need compliance, variable stiffness and safe interaction with humans.
• Lightweight manipulators for mobile platforms, where low inertia and a high payload-to-weight ratio matter.
• Research platforms for studying muscle synergy, co-contraction and stiffness modulation in robotic limbs.

Industry relevance: The hardware choices (mature DC motor technology, high-energy-density actuation, EtherCAT and UDP communication at up to 1000 Hz, 3D-printed Nylon HP3DHR-PA12 brackets) point toward manufacturable modular muscle units, while the model-free learning controller lowers the modeling burden for companies building compliant, tendon-driven manipulators.

Future Directions

• The paper notes that muscle layout configuration and actuation pathway design still require further optimization, and that uncertainties in muscle-to-bone contact and friction remain significant challenges for musculoskeletal platforms.
• The truncated experimental section begins to discuss the prototype's sensitivity to real-world conditions; the paper's own framing implies that closing the gap between simulated disturbance tolerance (up to roughly 25%) and prototype performance is an open task.
• Extending the DDILC approach beyond repetitive, finite-duration tasks — where iterative learning control applies — to non-repetitive or unstructured motions is an implied open question.
• Scaling the centralized artificial muscle architecture and the comparative benchmarking (skeletal similarity, muscle similarity, sensors, performance) to more complete humanoid musculoskeletal systems.

Target Audience

Robotics researchers and graduate students working on bio-inspired and musculoskeletal robot design, tendon-driven actuation, and learning-based or model-free nonlinear control. It is also relevant to engineers building compliant manipulators, prosthetics or humanoid platforms who are interested in combining mechanical compliance with data-driven iterative learning control. Readers without a background in nonlinear control theory or muscle biomechanics will find the modeling and control sections demanding.

Authors’ abstract

The human arm exhibits remarkable capabilities, including both explosive power and precision, which demonstrate dexterity, compliance, and robustness in unstructured environments. Developing robotic systems that emulate human-like operational characteristics through musculoskeletal structures has long been a research focus. In this study, we designed a novel lightweight tendon-driven musculoskeletal arm (LTDM-Arm), featuring a seven degree-of-freedom (DOF) skeletal joint system and a modularized artificial muscular system (MAMS) with 15 actuators. Additionally, we employed a Hilly-type muscle model and data-driven iterative learning control (DDILC) to learn and refine activation signals for repetitive tasks within a finite time frame. We validated the anti-interference capabilities of the musculoskeletal system through both simulations and experiments. The results show that the LTDM-Arm system can effectively achieve desired trajectory tracking tasks, even under load disturbances of 20 % in simulation and 15 % in experiments. This research lays the foundation for developing advanced robotic systems with human-like operational performance.