The human skin has several sensors with different properties and responses that are able to detect stimuli resulting from mechanical stimulations. Pressure sensors are the most important type of receptors for the exploration and manipulation of objects. In the last decades, smart tactile sensing based on different sensing techniques have been developed as their application in robotics and prosthetics is considered of huge interest, mainly driven by the prospect of autonomous and intelligent robots that can interact with the environment. However, regarding object properties estimation on robots, hardness detection is still a major limitation due to the lack of techniques to estimate it. Furthermore, finding processing methods that can interpret the measured information from multiple sensors and extract relevant information is a Challenging task. Moreover, embedding processing methods and machine learning algorithms in robotic applications to extract meaningful information such as object properties from tactile data is an ongoing challenge, which is controlled by the device constraints (power constraint, memory constraints, etc.), the computational complexity of the processing and machine learning algorithms, the application requirements (real-time operations, high prediction performance). In this dissertation, we focus on the design and implementation of pre-processing methods and machine learning algorithms to handle the aforementioned challenges for a tactile sensing system in robotic application. First, we propose a tactile sensing system for robotic application. Then we present efficient preprocessing and feature extraction methods for our tactile sensors. Then we propose a learning strategy to reduce the computational cost of our processing unit in object classification using sensorized Baxter robot. Finally, we present a real-time robotic tactile sensing system for hardness classification on a resource-constrained devices. The first study represents a further assessment of the sensing system that is based on the PVDF sensors and the interface electronics developed in our lab. In particular, first, it presents the development of a skin patch (multilayer structure) that allows us to use the sensors in several applications such as robotic hand/grippers. Second, it shows the characterization of the developed skin patch. Third, it validates the sensing system. Moreover, we designed a filter to remove noise and detect touch. The experimental assessment demonstrated that the developed skin patch and the interface electronics indeed can detect different touch patterns and stimulus waveforms. Moreover, the results of the experiments defined the frequency range of interest and the response of the system to realistic interactions with the sensing system to grasp and release events. In the next study, we presented an easy integration of our tactile sensing system into Baxter gripper. Computationally efficient pre-processing techniques were designed to filter the signal and extract relevant information from multiple sensor signals, in addition to feature extraction methods. These processing methods aim in turn to reduce also the computational complexity of machine learning algorithms utilized for object classification. The proposed system and processing strategy were evaluated on object classification application by integrating our system into the gripper and we collected data by grasping multiple objects. We further proposed a learning strategy to accomplish a trade-off between the generalization accuracy and the computational cost of the whole processing unit. The proposed pre-processing and feature extraction techniques together with the learning strategy have led to models with extremely low complexity and very high generalization accuracy. Moreover, the support vector machine achieved the best trade-off between accuracy and computational cost on tactile data from our sensors. Finally, we presented the development and implementation on the edge of a real–time tactile sensing system for hardness classification on Baxter robot based on machine and deep learning algorithms. We developed and implemented in plain C a set of functions that provide the fundamental layer functionalities of the Machine learning and Deep Learning models (ML and DL), along with the pre–processing methods to extract the features and normalize the data. The models can be deployed to any device that supports C code since it does not rely on any of the existing libraries. Shallow ML/DL algorithms for the deployment on resource–constrained devices are designed. To evaluate our work, we collected data by grasping objects of different hardness and shape. Two classification problems were addressed: 5 levels of hardness classified on the same objects’ shape, and 5 levels of hardness classified on two different objects’ shape. Furthermore, optimization techniques were employed. The models and pre–processing were implemented on a resource constrained device, where we assessed the performance of the system in terms of accuracy, memory footprint, time latency, and energy consumption. We achieved for both classification problems a real-time inference (< 0.08 ms), low power consumption (i.e., 3.35 μJ), extremely small models (i.e., 1576 Byte), and high accuracy (above 98%).
Signal and Information Processing Methods for Embedded Robotic Tactile Sensing Systems
AMIN, YOUSSEF
2023-03-17
Abstract
The human skin has several sensors with different properties and responses that are able to detect stimuli resulting from mechanical stimulations. Pressure sensors are the most important type of receptors for the exploration and manipulation of objects. In the last decades, smart tactile sensing based on different sensing techniques have been developed as their application in robotics and prosthetics is considered of huge interest, mainly driven by the prospect of autonomous and intelligent robots that can interact with the environment. However, regarding object properties estimation on robots, hardness detection is still a major limitation due to the lack of techniques to estimate it. Furthermore, finding processing methods that can interpret the measured information from multiple sensors and extract relevant information is a Challenging task. Moreover, embedding processing methods and machine learning algorithms in robotic applications to extract meaningful information such as object properties from tactile data is an ongoing challenge, which is controlled by the device constraints (power constraint, memory constraints, etc.), the computational complexity of the processing and machine learning algorithms, the application requirements (real-time operations, high prediction performance). In this dissertation, we focus on the design and implementation of pre-processing methods and machine learning algorithms to handle the aforementioned challenges for a tactile sensing system in robotic application. First, we propose a tactile sensing system for robotic application. Then we present efficient preprocessing and feature extraction methods for our tactile sensors. Then we propose a learning strategy to reduce the computational cost of our processing unit in object classification using sensorized Baxter robot. Finally, we present a real-time robotic tactile sensing system for hardness classification on a resource-constrained devices. The first study represents a further assessment of the sensing system that is based on the PVDF sensors and the interface electronics developed in our lab. In particular, first, it presents the development of a skin patch (multilayer structure) that allows us to use the sensors in several applications such as robotic hand/grippers. Second, it shows the characterization of the developed skin patch. Third, it validates the sensing system. Moreover, we designed a filter to remove noise and detect touch. The experimental assessment demonstrated that the developed skin patch and the interface electronics indeed can detect different touch patterns and stimulus waveforms. Moreover, the results of the experiments defined the frequency range of interest and the response of the system to realistic interactions with the sensing system to grasp and release events. In the next study, we presented an easy integration of our tactile sensing system into Baxter gripper. Computationally efficient pre-processing techniques were designed to filter the signal and extract relevant information from multiple sensor signals, in addition to feature extraction methods. These processing methods aim in turn to reduce also the computational complexity of machine learning algorithms utilized for object classification. The proposed system and processing strategy were evaluated on object classification application by integrating our system into the gripper and we collected data by grasping multiple objects. We further proposed a learning strategy to accomplish a trade-off between the generalization accuracy and the computational cost of the whole processing unit. The proposed pre-processing and feature extraction techniques together with the learning strategy have led to models with extremely low complexity and very high generalization accuracy. Moreover, the support vector machine achieved the best trade-off between accuracy and computational cost on tactile data from our sensors. Finally, we presented the development and implementation on the edge of a real–time tactile sensing system for hardness classification on Baxter robot based on machine and deep learning algorithms. We developed and implemented in plain C a set of functions that provide the fundamental layer functionalities of the Machine learning and Deep Learning models (ML and DL), along with the pre–processing methods to extract the features and normalize the data. The models can be deployed to any device that supports C code since it does not rely on any of the existing libraries. Shallow ML/DL algorithms for the deployment on resource–constrained devices are designed. To evaluate our work, we collected data by grasping objects of different hardness and shape. Two classification problems were addressed: 5 levels of hardness classified on the same objects’ shape, and 5 levels of hardness classified on two different objects’ shape. Furthermore, optimization techniques were employed. The models and pre–processing were implemented on a resource constrained device, where we assessed the performance of the system in terms of accuracy, memory footprint, time latency, and energy consumption. We achieved for both classification problems a real-time inference (< 0.08 ms), low power consumption (i.e., 3.35 μJ), extremely small models (i.e., 1576 Byte), and high accuracy (above 98%).File | Dimensione | Formato | |
---|---|---|---|
phdunige_4782119.pdf
accesso aperto
Tipologia:
Tesi di dottorato
Dimensione
14.62 MB
Formato
Adobe PDF
|
14.62 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.