Wenhui Yang

Shape recognition is essential in multifarious manipulations, such as object grasping, block building, and parts assembly, which most frequently relies on visual perception. However, it inevitably suffers from adverse impact of illumination intensity, visual distortion from different views, and frequent occlusions in dynamic scenes. To tackle this problem, the paper proposes recognizing different shapes via tactile exploration, mimicking the human manner of touch in dark environments. Based on our designed tactile fingertip, which comprises four sensing elements and a fingerprint structure, we employed the signal processing framework consisting of data preprocessing, feature pool establishment, selection of the most discriminative feature combination, and recognition model using a shallow neural network. Particularly, the feature pool was extended to comprise 386 multi-perspective features spanning different domains and scales. When the nine features with the highest weights were selected using Neighborhood Component Analysis, our extreme learning machine-based recognition model achieved an accuracy as high as 99.72%. To contextualize this performance, a comprehensive benchmark was conducted against several state-of-the-art deep learning paradigms, including end-to-end and transfer learning models. The analysis reveals that for data-constrained and resource-sensitive tactile sensing applications, our proposed and interpretable feature engineering approach not only achieves superior accuracy but also offers significant advantages in computational efficiency. This work validates the practical value of tailored feature engineering and sheds light on more efficient utilization of tactile signals in real-world robotic recognition tasks.

Robotic assembly of electrical connectors enables the automation of high-efficiency production of electronic products. A rigid gripper is adopted as the end-effector by the majority of existing works with a force–torque sensor installed at the wrist, which suffers from very limited perception capability of the manipulated objects. Moreover, the grasping and movement actions, as well as the inconsistency between the robot base and the end-effector frame, tend to result in angular misalignment, usually leading to assembly failure. Bio-inspired by the human finger, we designed a tactile finger in this paper with three characteristics: (1) Compliance: A soft ‘skin’ layer provides passive compliance for plenty of manipulation actions, thus increasing the tolerance for alignment errors. (2) Tactile Perception: Two types of sensing elements are embedded into the soft skin to tactilely sense the involved contact status. (3) Enhanced manipulation force: A rigid fingernail is designed to enhance the manipulation force and enable potential delicate operations. Moreover, a tactile-based alignment algorithm is proposed to search for the optimal orientation angle about the z axis. In the application of U-disk insertion, the three characteristics are validated and a success rate of 100% is achieved, whose generalization capability is also validated through the assembly of three types of electrical connectors.

To date, numerous tactile sensors and algorithms have been developed to tackle various perception problems. However, excessive endeavors are devoted to the multifarious definition, extraction, and analysis of hand-crafted features in order to improve the perception accuracy. To address this problem, in this paper, we designed a tactile finger containing four sensing elements to perceive both dynamic and static stimuli, and meanwhile proposed a novel signal processing pipeline. The pipeline mainly consisted of time-series signals conversion, automatic deep features extractor, and a shallow recognition model. When the tactile finger was applied to explore 16 surfaces based on a robotic platform, the four-channel signals were converted and concatenated into a time-frequency image via continuous wavelet transform. A deep feature extraction network was constructed based on a pretrained deep learning (DL) model, Resnet101, to extract the required features, which acted as high-level representations of the most discriminative components from the tactile images. Finally, these features were fed into a shallow machine learning model, i.e., extreme learning machine, achieving an accuracy as high as 92.38%. In such a manner, the powerful learning capability of DL models was transferred to the new recognition model directly while the tedious feature extraction procedures were alleviated. Besides, several relevant issues, such as the layer depth, the DL model type, and the shallow recognition model, are addressed and discussed to reveal their influences on the performance.