Introduction
Industrial robotics has long excelled in controlled environments requiring high precision and heavy lifting. However, these systems often struggle with the dexterity and compliance required for unstructured or unpredictable human-centric settings. Replicating the human arm's capabilities presents two primary challenges: the mechanical complexity of rigid-flexible coupling between bones and muscles, and the control redundancy inherent in over-actuated musculoskeletal systems. This study introduces a lightweight tendon-driven musculoskeletal arm (LTDM-Arm) and a data-driven iterative learning control (DDILC) algorithm designed to overcome these hurdles without requiring an exhaustive mathematical model of the system.
Hardware Design and Modeling
The LTDM-Arm is a 7-degree-of-freedom (DOF) robotic system modeled after human anatomy, including the shoulder blade, humerus, ulna, and radius. It utilizes a Modularized Artificial Muscular System (MAMS) consisting of 15 muscle units driven by DC motors and high-strength polyester cables. These cables act as tendons, routed through lubricated PTFE sheaths to minimize friction.
To achieve biological realism, the researchers integrated a Hill-type muscle model. This model accounts for both the active contraction of muscle fibers ($F^{ce}$) and the passive elasticity of tendons ($F^{pe}$), effectively simulating the non-linear dynamics of human movement by incorporating length and velocity dependencies.
Control Strategy: Data-Driven Iterative Learning
The core of the system's intelligence is the Data-Driven Iterative Learning Control (DDILC) algorithm. Unlike traditional control methods that rely on rigid physical models, DDILC leverages input/output data from previous task repetitions to refine control signals. By focusing on how errors change between iterations, the system "learns" the necessary muscle activations.
This approach is particularly effective for managing the "over-actuated" nature of the arm, where multiple muscles (antagonistic pairs) work in synergy to stabilize joints and smooth trajectory errors. Furthermore, the muscle model acts as a natural low-pass filter, dampening high-frequency control noise through inherent activation delays and tendon elasticity.
Results and Discussion
The performance of the LTDM-Arm was evaluated through numerical simulations in Mujoco-Python and physical experiments involving sine-wave trajectory tracking.
- High Precision: In simulation, the average trajectory error was reduced to 1.38 mm (0.14%). On the physical prototype, the error converged to 3.8 mm (0.4%) after 40 iterations.
- Robustness to Disturbance: The system maintained stability under significant external disturbances, successfully tracking trajectories with load variations of up to 20% in simulation and 15% in physical trials.
- Efficiency: Compared to baseline methods like Computed Muscle Control (CMC) and Nonlinear-PID, the DDILC approach achieved error reduction rates between 57.8% and 74.13%, proving its superior ability to manage the nonlinearities of cable-driven systems.
While the results demonstrate high adaptability, the physical prototype showed increased sensitivity to manufacturing tolerances and joint friction compared to simulations, leading to slightly higher real-world error margins.
Conclusion
The LTDM-Arm and DDILC algorithm provide a robust framework for bio-inspired robotic control. By combining human-like morphology with a data-driven learning approach, the system achieves high precision and stability in repetitive tasks. Future research will focus on integrating multi-modal sensory data—such as vision and touch—to enable safer human-robot interaction and the execution of complex, non-repetitive assembly tasks in unstructured environments.
Summary updated