Real-time Puncture Detection and Recovery for Pneumatic Soft Actuators

I. INTRODUCTION

• Punctures and leaks can impair performance and safety, especially in damage-susceptible soft materials, motivating detection and recovery strategies.The authors identify biodegradable materials such as gelatin and hydrogel as particularly susceptible to damage.Source: §I · p. 1
• Over 96% accuracy is reported for localizing punctured chambers from time-series data collected by a single IMU.The algorithm also estimates severity on a 0-to-1 scale.Source: §I · p. 1
• The multi-chamber bending actuator uses deformation from intact chambers to continue operating after a chamber failure.The actuator chambers can be selectively inflated.Source: §I · p. 2
• The paper presents a puncture-detection algorithm using only one compact, low-power, 6-degree-of-freedom onboard IMU.The authors frame this as suitable for untethered field robots with payload constraints.Source: §I · p. 2
• The contributions include chamber localization and severity scoring, a perturbation and recovery scheme, and an IMU dataset of punctures across chambers.The recovery scheme is intended to retain actuator tip-force after puncture.Source: §I · p. 2

A. Puncture Detection for Soft Actuators

• Anomaly detection in soft robots remains largely unexplored, with deformability and compliance complicating sensing, modeling, and experimental repeatability.The section groups anomaly-detection methods as supervised, semi-supervised, or unsupervised.Source: A. Puncture Detection for Soft Actuators · p. 2
• Unsupervised approaches avoid labeled training data, but lack of ground truth can make evaluation difficult and contribute to false positives in complex settings.The section cites clustering methods including k-means and distance-based clustering.Source: A. Puncture Detection for Soft Actuators · p. 2
• Semi-supervised detectors can learn normal-only data; autoencoders flag unfamiliar patterns through high reconstruction error.Examples discussed include OC-SVMs and isolation forests.Source: A. Puncture Detection for Soft Actuators · p. 2
• The paper uses a fully connected autoencoder for puncture localization and a single tanh-activated perceptron for severity scoring.The authors cite simplicity, interpretability, and efficiency on smaller datasets as reasons for the autoencoder choice.Source: A. Puncture Detection for Soft Actuators · p. 2

B. Sensing for failure detection

• Embedded strain, capacitive, pressure, optical-fiber, and liquid-metal sensors can provide real-time measurements but may add complexity or compromise softness.The section gives liquid-metal droplets embedded in silicone as an example of damage detection.Source: B. Sensing for failure detection · p. 2
• IMUs offer compact, low-cost sensing for control and state estimation without compromising actuator compliance.The section notes prior use of IMUs in soft robots.Source: B. Sensing for failure detection · p. 2
• Pressure sensing in every chamber can raise cost and integration complexity, while soft-material nonlinearity and hysteresis can limit accurate dynamic measurements.Frequent pressure drops may indicate a puncture.Source: B. Sensing for failure detection · pp. 2–3
• The paper selects IMUs as its only sensing modality, using deviations in expected motion to indicate puncture-related changes in pressure or actuation efficiency.Source: B. Sensing for failure detection · p. 3

A. Multi-Chamber Soft Actuator Design

• Each chamber in the pneumatic bending actuator can be inflated or deflated individually, allowing intact chambers to continue operation after failures.The design aims to prevent total actuator failure if one or more chambers are punctured.Source: A. Multi-Chamber Soft Actuator Design · p. 3

B. Experimental Test Bed Setup

• The suspended testbed combines an actuator mount, pneumatic connections, control electronics, and a portable pump.The suspension is intended to reduce uncertainty from friction and surface interactions.Source: B. Experimental Test Bed Setup · p. 3

1) Data Collection for Puncture Localization:

• Data collection perturbed chambers in normal and puncture cases while recording 6-axis IMU data; Dataset 1 contained 1,388 samples.Dataset 1 had 682 working and 706 fail samples and was used for training.Source: 1) Data Collection for Puncture… · p. 3
• Datasets 2 and 3 tested robustness under multi-chamber failure and tip-contact damping, variations absent from training data.Dataset 2 tested chamber 4 with chambers 3 and 5 failed; Dataset 3 involved consistent actuator-tip contact with a stationary object.Source: 1) Data Collection for Puncture… · p. 3
• The perturbation used a 20 ms inflation pulse, a 500 ms delay, and a 20 ms deflation pulse, with timing chosen empirically from valve response times.Source: 1) Data Collection for Puncture… · pp. 3–4

• Four deflation pulses were used after the perturbation, with 50 samples collected for each of four puncture diameters and the no-puncture case.The tested circular punctures were 0.51, 1.06, 1.30, and 2.00 mm.Source: 2) Data Collection for Severity Score · p. 4
• A delay between inflation and deflation separates their responses, reduces vibration noise, and allows leakage to manifest.The perturbation keeps the actuator pose close to its initial configuration; the deflation response differed markedly between working and punctured chambers.Source: 2) Data Collection for Severity Score · p. 4

D. Feature Engineering

• The six-axis IMU yields 60 statistical and impulsive features per time series, ranked by t-test scores, with the top 20 selected for localization and severity modeling.The feature set includes mean, standard deviation, RMS, shape factor, kurtosis, skewness, crest factor, impulse factor, clearance factor, and peak value.Source: D. Feature Engineering · p. 4

E. Puncture Localization and Severity Score Algorithm

• IMU data are segmented into seven segments, each corresponding to exactly one chamber after chambers are perturbed one by one.Source: E. Puncture Localization and Severity… · p. 4
• One-class SVMs, isolation forests, and FCN autoencoders are trained only on working data to detect puncture-related deviations.The autoencoder flags failure when reconstruction error exceeds a predefined threshold.Source: E. Puncture Localization and Severity… · p. 4
• Severity is modeled as a continuous score from 0 to 1 using a single-layer perceptron with tanh activation to capture nonlinear pressure-drop relationships.Source: E. Puncture Localization and Severity… · p. 4

A. Localization of punctured chamber

• 96.85% accuracy was achieved by the FCN autoencoder, with 0.97 precision and 0.96 recall on puncture detection.Mean reconstruction error was 20.42 for fail data and 1.76 for working data; the detection summary is shown in Fig. 5.Source: A. Localization of punctured chamber · p. 5, Figure 5 · p. 5
• 99% accuracy on Dataset 2 was maintained without fine-tuning the model trained on Dataset 1, using threshold 0.4.The passage attributes this high accuracy to data collected from just one chamber.Source: A. Localization of punctured chamber · p. 5, Figure 5 · p. 5
• On Dataset 3, contact-induced damping reduced accuracy below 80% at the original threshold; threshold 1.6 raised it to 91%.Source: A. Localization of punctured chamber · p. 5, Figure 5 · p. 5
• On Dataset 1, one-class SVM achieved 99.03% accuracy, while isolation forest achieved 89.59%; both fell below 80% accuracy on Datasets 2 and 3.The authors associate the generalization drop with signal changes, including damping from adjacent chamber failures or environmental interactions absent from training.Source: A. Localization of punctured chamber · p. 5

• Minor punctures may leave enough functionality for continued operation, whereas severe punctures may warrant excluding the chamber and using system redundancy.Higher scores indicate more severe punctures and rapid pressure loss.Source: B. Puncture Severity Score · p. 6
• The severity score α is labeled from t0, the time to return to initial pressure after the same inflation cycle; no puncture maps to 0 and a 2.00 mm puncture maps to 1.The score uses the tested minimum and maximum t0 values, since smaller t0 is not linearly more severe.Source: B. Puncture Severity Score · p. 6
• The IMU-feature regression model uses a single-layer perceptron with tanh activation and achieved a mean square error of 0.0015.The first two principal components show clustered samples corresponding to different puncture severities.Source: B. Puncture Severity Score · p. 6

C. Puncture Failure Recovery

• The candidate recovery strategy treats actuator tip force as the control variable and compensates for errors caused by high-severity punctures.During normal operation, all chambers are pressurized equally to achieve the desired tip force; failures can prevent that force under the same input.Source: C. Puncture Failure Recovery · p. 7
• A threshold on deviation from desired tip force triggers puncture detection, with a short delay to avoid immediate post-failure transients.The passage identifies failures in chambers 4 and 6 as examples.Source: C. Puncture Failure Recovery · p. 7
• Chambers with severity scores below 0.3 may remain active, while intact chambers are inflated equally toward the desired tip force.Depending on the number of failed chambers, recovery may not be possible; its scope requires further investigation.Source: C. Puncture Failure Recovery · p. 7

V. DISCUSSION

• 98% localization accuracy was achieved by fully connected autoencoders even under out-of-distribution scenarios, while OC-SVMs and isolation forests degraded under unseen external perturbations.The discussion contrasts the statistical methods’ simplicity and interpretability with the stronger resilience of autoencoder-based anomaly detection.Source: §V · p. 7
• Further validation is needed across soft-robot designs and external disturbances; feature-extraction models maintained high accuracy under static loads, but dynamic loads may obscure puncture signatures.For multi-segment robots, the discussion proposes perturbing each segment and selecting sensor placement to capture failure-related motion signatures across segments.Source: §V · p. 7
• The framework has been validated only on Ecoflex 00-50; evaluation across other elastomers and simultaneous punctures in different actuator designs remains future work.The authors also plan to test alternative silicone materials with diverse mechanical properties.Source: §V · p. 7
• Extending the method to biodegradable, mechanically fragile materials such as gelatin-based actuators is planned future work.The authors describe this as an extension for sustainable alternatives to conventional soft robots.Source: §V · p. 8

VI. CONCLUSIONS

• The in-situ method uses onboard IMU data collected through chamber perturbation for puncture localization and severity estimation.Localization uses autoencoder-based models, while severity is estimated with non-linear regression, supporting failure-informed control without external instrumentation.Source: §VI · p. 8
• A candidate recovery strategy for multi-chamber failures illustrates the framework’s practical utility.Source: §VI · p. 8
• The authors describe the method as scalable and cost-effective, with potential adaptation to a wide range of pneumatic actuator designs and materials.They link this potential to reliance on minimal sensing infrastructure.Source: §VI · p. 8