A Smart Capacitive Sensor Skin with Embedded Data Quality Indication for Enhanced Safety in Human-Robot Interaction

Christoph Scholl, Andreas Tobola, Klaus Ludwig, Dario Zanca, Bjoern M. Eskofier

TL;DR

This paper introduces a novel capacitive sensor skin that not only detects object proximity and angle but also embeds a real-time data quality monitoring system. This integrated approach significantly enhances the safety of human-robot interactions by flagging unreliable sensor data, leading to improved prediction accuracy and system reliability.

What Problem Does This Paper Address?

Human-robot interaction (HRI) is increasingly reliant on smart sensors to ensure safety. While advancements in machine learning have made sensors more powerful, a critical challenge remains: ensuring the quality of the data these sensors produce, especially in safety-critical applications. Poor data quality can lead to incorrect decisions, compromising the safety of both humans and robots. Existing sensor systems often neglect this aspect, or their data quality monitoring solutions are computationally too demanding for embedded systems. This paper addresses the need for an efficient, embedded solution for real-time data quality monitoring in capacitive sensor systems used for HRI.

Methods

The core of this work is a capacitive sensor skin designed to detect conductive objects within a 200 mm range, estimating both distance and angle. The system employs a two-pronged approach:

  • Object Detection: A fully connected neural network is used to infer the position and angle of nearby objects based on sensor readings.
  • Data Quality Monitoring: A dedicated layer, implemented using a one-class Support Vector Machine (SVM), monitors data quality in real-time. This SVM acts as an out-of-distribution (OOD) detector, identifying samples that deviate from the expected normal data distribution.

The system also incorporates temperature and drift compensation modules to improve the robustness of the sensor readings. The algorithms are designed to be lightweight and computationally efficient to run on embedded microcontrollers with limited resources.

Datasets

The paper describes the acquisition of several datasets using a custom-built test bench.

  • Training and Validation: Data set 1 was used for training and validation of both the neural network and the data quality indication algorithms. This dataset was filtered to exclude measurements beyond the 200 mm detection range, shuffled, and split into training, testing, and validation sets.
  • Testing and Evaluation: Data sets 2 through 8 were used to test algorithm performance, specifically focusing on data quality indication based on OOD detection and the accuracy of distance and angle predictions. These datasets were acquired over various days to account for different environmental conditions.

Key Insights and Ideas

  • Data Quality is Crucial for Safety: The paper strongly emphasizes that in safety-critical HRI, simply having accurate sensor readings isn't enough; knowing when those readings are unreliable is equally, if not more, important.
  • Embedded OOD Detection for Efficiency: Using a one-class SVM for out-of-distribution detection on the sensor itself is a key innovation. This allows for real-time data quality assessment without needing to offload heavy computation to a separate system, making it suitable for resource-constrained embedded devices.
  • Trade-offs in Complexity: The study highlights the inherent trade-off between the complexity of data quality algorithms (like different SVM variants) and their performance metrics (accuracy vs. computational cost). More complex models might offer better accuracy but are too computationally expensive for embedded deployment.
  • Significant Performance Improvement: Integrating data quality indication demonstrably improves the overall performance of the sensor system, reducing errors and increasing reliability.

Key Results

  • Distance Prediction Improvement: Without data quality indication, the sensor achieved a Mean Absolute Error (MAE) of 11.64 mm. With data quality indication using the SVM-829 algorithm, the MAE was reduced to 7.4 mm, a significant improvement.
  • Angle Prediction Improvement: The average angular error was reduced from 15.6° to 14.1° when data quality was considered using SVM-829.
  • Computational Efficiency: The neural network had an execution time of 0.14 ms, and the one-class SVM decision function took 1.54 ms, both well within the 5 ms time budget for processing a single sample.
  • Temperature Compensation Effectiveness: The temperature compensation module reduced temperature coefficients by an average of 83.5%, making the system more stable.

Limitations and Future Work

  • Time Dependency Ignored: The current system and algorithms do not explicitly consider the time dependency of input signals for either the neural network or the OOD detection.
  • Future Work - Time Series Analysis: The authors suggest exploring more complex data-driven approaches like Long Short-Term Memory (LSTM) or Recurrent Neural Networks (RNN) to incorporate time-series data. However, they acknowledge this might increase system latency.
  • Future Work - Knowledge-Based Approaches: Investigating knowledge-based methods for data quality indication is another avenue for future research.
  • Future Work - Broader Error Handling: The current OOD detection doesn't cover all types of sensor performance errors; future work should address additional data quality-related parameters.
  • Generalizability: The proposed system architecture and data quality module are presented as adaptable to various use cases beyond HRI to enhance sensor reliability.

Summary updated

Abstract

Smart sensors are an integral part of the Fourth Industrial Revolution and are widely used to add safety measures to human–robot interaction applications. With the advancement of machine learning methods in resource-constrained environments, smart sensor systems have become increasingly powerful. As more data-driven approaches are deployed on the sensors, it is of growing importance to monitor data quality at all times of system operation. We introduce a smart capacitive sensor system with an embedded data quality monitoring algorithm to enhance the safety of human–robot interaction scenarios. The smart capacitive skin sensor is capable of detecting the distance and angle of objects nearby by utilizing consumer-grade sensor electronics. To further acknowledge the safety aspect of the sensor, a dedicated layer to monitor data quality in real-time is added to the embedded software of the sensor. Two learning algorithms are used to implement the sensor functionality: (1) a fully connected neural network to infer the position and angle of objects nearby and (2) a one-class SVM to account for the data quality assessment based on out-of-distribution detection. We show that the sensor performs well under normal operating conditions within a range of 200 mm and also detects abnormal operating conditions in terms of poor data quality successfully. A mean absolute distance error of 11.6mm was achieved without data quality indication. The overall performance of the sensor system could be further improved to 7.5mm by monitoring the data quality, adding an additional layer of safety for human–robot interaction.
BibTeX
@inproceedings{scholl2021smart,
  author = {Christoph Scholl and Andreas Tobola and Klaus Ludwig and Dario Zanca and Bjoern M. Eskofier},
  title = {{A Smart Capacitive Sensor Skin with Embedded Data Quality Indication for Enhanced Safety in Human-Robot Interaction}},
  booktitle = {Sensors},
  year = {2021},
  url = {https://openreview.net/forum?id=OhZoYOEDFSq},
}

Loading related content...

Figure