Damage detection in soft robotics presents a persistent challenge, particularly in industrial applications where repetitive tasks can turn minor cuts or fatigue into catastrophic failures. Traditional approaches often embed sensor networks within the soft body, mimicking biological pain perception through conductive pathways, strain-sensitive materials, or pressure-sensing chambers. While effective, these methods require complex fabrication, additional wiring, and can compromise the compliance that makes soft robots advantageous. They are also difficult to retrofit onto existing systems and may suffer from reduced durability under high-cycle industrial use.
A recent study introduces a non-invasive, data-driven alternative for detecting and localizing damage in soft grippers by analyzing changes in their motion dynamics. The core idea leverages the nonlinear nature of soft robotic systems: small morphological or material changes from damage can yield significant variations in force and torque feedback over time. Instead of embedding sensors into the soft structure, the method measures six-axis force/torque signals at the gripper’s mounting point and processes them with a bidirectional long short-term memory (biLSTM) neural network to classify damage states.
The experimental platform used a two-fingered Fin Ray gripper with interchangeable silicone fingers, each fabricated from Dragon Skin 20 silicone in 3D-printed molds. Damage was introduced in controlled configurations—partial cuts from either edge or complete cuts—at specific contact areas along the inner sides of the fingers. Each finger’s regions were indexed, and the damage type was labeled, yielding 42 damaged configurations plus an undamaged state, for a total of 43 classes.
To excite measurable dynamics, the gripper was mounted on a UR5 robotic manipulator operating in servoing mode. The setup oscillated the gripper at 4 Hz with a defined angular displacement, while varying its roll orientation through eight positions spaced at 45° intervals. An ATI Nano43 force/torque sensor captured raw strain gauge signals at 125 Hz, which were normalized but not filtered to preserve information-rich noise. Data segmentation used a sliding window of 125 points with a 10-point shift, producing over 54,000 training samples and nearly 11,000 test samples.
The biLSTM architecture comprised a single recurrent layer with 50 hidden units, chosen to balance performance and overfitting risk. Hyperbolic tangent and sigmoid functions served as activation functions for state and gate operations, respectively, with a dropout layer (p=0.5) for regularization. A fully connected layer mapped outputs to the 43 damage classes, followed by a softmax layer for probability normalization. Training employed stochastic gradient descent with the ADAM optimizer, a learning rate of 1e-3, and mini-batches of 600 samples.
Testing on unseen data achieved a 99% damage detection rate and over 97% localization accuracy. The “no-damage” class was correctly identified 89.4% of the time, but when treated as a binary detection problem, both precision and recall exceeded 0.99. Misclassifications were rare and often attributable to subtle physical effects, such as cut edges sticking together, which masked dynamic changes. Damage near the gripper tips proved harder to localize due to reduced inertia in the remaining structure.
Analysis of the input space revealed that gripper orientation significantly influenced classification accuracy. Horizontal orientations orthogonal to the oscillation direction (π/2 and 3π/2 roll angles) yielded the best results, approaching 96% accuracy even when used alone. Combining multiple less-informative orientations did not outperform these optimal single orientations. Time-series length also played a role: as few as 25 data points (half an oscillation period) provided reasonable localization, while 75 points (about two periods) unlocked the model’s full potential.
By avoiding embedded sensing elements, this approach offers a retrofit-friendly, cost-effective solution for monitoring passive soft grippers, particularly cable-driven designs where pneumatic pressure-based diagnostics are unavailable. The method’s reliance on external force/torque sensing and controlled excitation makes it adaptable to other soft actuator types, provided their dynamics can be consistently stimulated.
Future work aims to extend the technique to arbitrary operational motions, eliminating the need for predefined oscillation patterns, and to explore regression models for continuous damage characterization. Additional directions include applying the method to other actuation schemes—tendon-driven, pneumatic, or hydraulic—and investigating its potential for non-invasive material property identification through nonlinear dynamic analysis.