Neural Network-Based Self-Sensing for Shape Memory Alloy-Driven Continuum Robots Control

Continuum robots have gained increasing attention in fields such as industrial automation, medical robotics, and security applications, primarily due to their structural flexibility. Among various actuation technologies, shape memory alloy (SMA) wire actuators stand out for their high energy density, compactness, and dual functionality as both actuators and sensors. During actuation, variations in the electrical resistance of SMA wires can be measured in a self-sensing manner, eliminating the need for external sensors or vision systems. These measurements can be used to estimate the wire length, enabling real-time configuration estimation while significantly reducing system complexity, cost, and size.

Traditionally, the interpretation of self-sensing data in SMA-driven continuum robots has relied on complex kinematic models. While effective, such methods typically demand considerable computational resources and processing time, thereby limiting their applicability in real-time control scenarios. This study investigates the use of neural networks as an alternative approach to self-sensing, with the objective of reducing computational complexity. The proposed method enables more efficient, compact, and responsive robotic systems by minimizing dependence on traditional kinematic modeling and intensive data processing.

To achieve this, a motion data collection setup was developed to validate a kinematic model describing the robot’s deformation. This validated model was then used to generate training data for the neural network, using self-sensing signals as input and the model’s predicted configuration as output. Finally, the trained network was evaluated using the same setup employed for kinematic model validation.