TL;DR
This paper presents the first computational framework for generating feasible motion trajectories for soft robots actuated by shape-memory alloys (SMAs). By developing a simplified dynamics model and using direct collocation optimization, they demonstrate the ability to accurately reproduce complex motions on a hardware prototype, paving the way for more advanced soft robotic applications.
What Problem Does This Paper Address?
Soft robots offer unique advantages like biomimicry and safe human interaction, but their practical use is hindered by the difficulty in generating complex, precise motions. This is particularly true for soft robots actuated by thermally-responsive materials like SMAs, which are challenging to model and control due to their inherent nonlinearities and thermal dynamics. Existing methods often lack dynamic feasibility guarantees or require computationally intensive models. This paper aims to bridge this gap by providing a computationally tractable method for generating feasible trajectories for SMA-actuated soft robots.
Methods
The authors propose a three-part approach:
- Simplified Dynamics Model: They approximate the soft limb as a discretized rigid manipulator with torsional springs. The actuation is modeled as joint torques proportional to SMA wire temperature, derived from a simplified SMA constitutive model and Joule heating principles. This model is designed for low computational cost.
- Model Calibration: The simplified dynamics model is calibrated using experimental data from a custom hardware platform. This involves identifying parameters for passive dynamics (spring constant, damping) and active dynamics (thermal actuator constants, force coefficients).
- Trajectory Optimization: A direct collocation optimization routine is employed. This method finds feasible state (joint angles, temperatures) and input (PWM duty cycles) trajectories that minimize a cost function (e.g., deviation from a reference trajectory) while satisfying constraints such as actuator temperature limits and warm-up periods.
Datasets
The paper does not explicitly mention training datasets. Instead, it focuses on collecting experimental data from a custom hardware platform for model calibration and validation. This includes:
- Static deflection data: Used to calibrate the linear torsional spring constant.
- Passive oscillation data: Used to calibrate the damping constant.
- Actuated data (single and dual actuator): Used to calibrate thermal actuator constants and force coefficients.
- Hardware tests: Three experiments (one hand-derived reference, two "teach and repeat") were performed to validate the optimized trajectories against hardware execution.
Key Insights and Ideas
- Temperature as a proxy for SMA actuation: The paper demonstrates that SMA wire temperature, rather than complex stress-strain relationships, can be effectively used to model and control the actuation of soft robots, significantly simplifying the dynamics.
- Rigid manipulator approximation for soft robots: Using a discretized rigid manipulator model, despite its apparent simplicity, proves sufficient for capturing the essential dynamics of the soft limb for trajectory generation purposes.
- Feasibility guarantees through optimization: Direct collocation optimization provides a crucial advantage by ensuring that the generated trajectories are dynamically feasible a priori, a significant challenge for many control techniques in soft robotics.
- "Teach and Repeat" for soft robots: The ability to optimize and execute "teach and repeat" motions opens up possibilities for intuitive robot programming and adaptation in soft robotic systems.
Key Results
- The calibrated model accurately predicts the limb's motion in open-loop hardware tests, with simulations qualitatively aligning with hardware rollouts.
- The trajectory optimization framework successfully generated feasible trajectories for various motions, including step inputs, decaying sinusoids, and "teach and repeat" motions.
- Tracking errors for the optimized trajectories were relatively small, with a median absolute error of 2.33° to 5.09° across the three hardware tests, which is considered reasonable given the sensor's capabilities and the complexity of the system.
Limitations and Future Work
Limitations:
- Planar and single-segment: The methodology is currently limited to planar soft robots and has only been verified on a single-segment limb with two actuators.
- Model approximations: Simplifications in the SMA model (e.g., neglecting phase transformation dynamics) and the temperature measurement lag introduce modeling errors.
- Temperature constraints: The approach is limited for long-term operation due to the need for warm-up periods and potential overheating of SMAs.
- Stiffness increase: As SMAs heat up, they become stiffer, which can sacrifice compliance and increase power consumption.
Future Work:
- 3D and multi-limbed robots: Extending the framework to handle 3D motion and multi-limbed robots.
- Locomotion: Applying the approach to generate walking gaits for robots like the one described in reference [8].
- Closed-loop control: Developing strategies to address model mismatch and unmodeled environments, potentially using techniques like iterative learning control or model-predictive control.
- Improved hardware calibration and SMA modeling: Enhancing the accuracy of calibration procedures and incorporating more sophisticated SMA models.
Summary updated