Robust Control of a Multi-Axis Shape Memory Alloy-Driven Soft Manipulator

Zach J. Patterson, Andrew P. Sabelhaus, Carmel MajidiIEEE RA-L 2022 · DOI

TL;DR

This paper presents a novel robust feedback control framework for multi-axis, 3D-moving soft robotic manipulators driven by shape memory alloy (SMA) actuators. By combining a simplified static model with advanced control techniques like Singular Value Decomposition (SVD) and anti-windup, they achieve precise trajectory tracking, overcoming challenges posed by actuator saturation and complex multi-axis dynamics.

What Problem Does This Paper Address?

Controlling soft robots, especially those capable of complex, multi-axis 3D motion, is a significant challenge. Traditional robots are rigid, making their dynamics easier to model and control. Soft robots, with their inherent flexibility and continuously deforming bodies, present a much harder modeling problem. This difficulty is amplified when using actuators like Shape Memory Alloys (SMAs), whose thermal dynamics and saturation limits are complex to manage. Existing control methods for soft robots are often limited to planar motion, simplified actuator models, or lack rigorous stability proofs. This paper tackles the dual challenge of achieving precise 3D motion control for SMA-driven soft manipulators while accounting for actuator saturation and unmodeled dynamics.

Methods

The authors propose a model-based control approach that leverages a simplified static beam bending model to represent the soft limb as a Linear Time-Invariant (LTI) system. This model, while not capturing full dynamics, allows for the application of robust control theory.

Key components of their control architecture include:

  • Singular Value Decomposition (SVD) Control: This technique is used as a pre- and post-compensator to decouple the multi-axis motion. By analyzing the singular values of the plant's transfer function, they can effectively "shape" the plant and design controllers that address interactions between different axes, making the system easier to control.
  • Diagonal PI Controller: A proportional-integral (PI) controller is used as the core feedback element. The PI controller is chosen for its broad effectiveness and ease of use.
  • Hanus Anti-Windup Conditioner: To address actuator saturation (where the controller commands a value beyond the actuator's physical limits), a Hanus conditioner is implemented. This mechanism prevents integral terms in the controller from accumulating excessive error during saturation, which can destabilize the system.
  • Input Saturation Adjustment: An additional adjustment is made to the control input to maintain the direction of the commanded input even when saturation occurs, further enhancing robustness.

The stability of the closed-loop system, including the saturation nonlinearity, is proven using a framework that accounts for both unmodeled dynamics and actuator saturation.

Datasets

Not discussed in the paper. The evaluation is performed on a custom-built hardware prototype.

Key Insights and Ideas

  • Simplified Models Can Be Powerful: The paper demonstrates that even a simplified static model of the soft limb, when combined with robust control techniques, can yield excellent performance for complex 3D motion. This suggests that full dynamic modeling might not always be necessary for achieving precise control in soft robotics.
  • Decoupling is Key for MIMO Control: Using SVD to decouple the multi-axis motion is crucial for effectively controlling systems with interacting inputs and outputs, like this soft manipulator.
  • Robust Control for Actuator Constraints: The integration of anti-windup mechanisms is essential for practical control of actuators with saturation limits, preventing performance degradation and instability.
  • Hardware Validation is Crucial: The successful implementation and testing on a physical SMA-driven soft robot limb provide strong evidence for the effectiveness of their proposed control framework.

Key Results

The controller demonstrated low-error trajectory tracking, with errors consistently within 5° to 10° for both yaw and pitch angles, even for fast commands and large deformations. This performance is comparable to the inherent uncertainty of their bending sensor.

  • Step Response: Higher proportional gains (Kp = 2.0) resulted in significantly lower steady-state errors (around 2-3°) compared to lower gains (Kp = 0.5, around 3-4°).
  • Trajectory Tracking: The controller successfully tracked trajectories, with performance degrading slightly at higher speeds. The higher gain (Kp = 2.0) consistently achieved better tracking.
  • Robustness: The stability proof and hardware validation indicate that the controller is robust to unmodeled dynamics and actuator saturation.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Robust Performance vs. Stability: The current stability analysis is conservative. Future work could focus on achieving robust performance rather than just stability, potentially using less conservative stability proofs that account more precisely for actuator saturation.
  • Model Accuracy: The Euler-Bernoulli beam model assumes small deflections. For larger deformations, a more advanced model (e.g., Timoshenko) might be necessary. Explicitly modeling and controlling velocity dynamics is also a future direction.
  • Artificial Stiffness: The PI controller, while effective, introduces artificial stiffness. Future work aims to reduce this stiffness by incorporating more substantial feedforward elements and internal actuator state sensing for improved disturbance rejection.
  • Generalization: While the framework is presented as a general solution, further work could explore its application to more complex soft robots, such as multi-link serial chains or legged robots.

Summary updated

BibTeX
@article{patterson2022robust,
  author = {Zach J. Patterson and Andrew P. Sabelhaus and Carmel Majidi},
  title = {{Robust Control of a Multi-Axis Shape Memory Alloy-Driven Soft Manipulator}},
  journal = {IEEE RA-L},
  year = {2022},
  doi = {10.1109/lra.2022.3143256},
}
Figure