International Journal of Humanoid Robotics

Although deep learning models are state-of-the-art models in audio classification, they fall short when applied in developmental robotic settings and human–robot interaction (HRI). The major drawback is that deep learning relies on supervised training with a large amount of data and annotations. In contrast, developmental learning strategies in human–robot interaction often deal with small-scale data acquired from HRI experiments and require the incremental addition of novel classes. Alternatively, shallow learning architectures that enable fast and yet robust learning are provided by simple distance metric-based learning and neural architectures implementing the reservoir computing paradigm. Similarly, continual learning algorithms receive more attention in the last years as they can integrate stable perceptual feature extraction using pre-trained deep learning models with open-set classification. As our research centers around reenacting the incremental learning of audio cues, we conducted a study on environmental sound classification using the iCaRL as well as the GDumb continual learning algorithms in comparison with a popular classifier in this domain, the kNN classifier, as well as employing an Echo State Network. We contrast our results with those obtained from a VGGish network that serves here as the performance upper bound that allows us to quantify the performance differences and to discuss current issues with continual learning in the audio domain. As only little is known about using shallow models or continual learning in the audio domain, we pass on additional techniques like data augmentation and create a simple experimental pipeline that is easy to reproduce. Although our selected algorithms are partially inferior in performance compared to the upper bound, our evaluation on three environmental sound datasets shows promising performance using continual learning for a subset of the DCASE2019 challenge dataset and the ESC10 dataset. As we do not address benchmarking in this paper, our study provides a good foundation for further research and computational improvements on shallow and continual learning models for robotic applications in the audio domain.

Gait recognition is one of the key technologies for exoskeleton robot control. The recognition accuracy and robustness of existing gait recognition methods cannot well meet the needs of real-time control. There is still room for improvement in fine-grained gait recognition. In this regard, this paper proposes a gait recognition method based on the MiniRocket and inertial measurement units. In this paper, a human lower limb posture information collection device is developed to collect ten kinds of gait data of human lower limbs (walking, standing, running, backing off, going upstairs, going downstairs, going uphill, going downhill, stand at ease and squat). The MiniRocket algorithm was used to build a human gait recognition model, and the effects of algorithm parameters and the size of the window and shift on the performance of gait recognition were discussed, and user-independent experiments and user-dependent experiments were carried out, respectively, and compared with four algorithms of TST, TCN, RNN and LSTM. The experimental results show that the MiniRocket algorithm has an average recognition accuracy of 94.87% and 97.67% in the user-independent experiment and the user-dependent experiment, which is better than the other four algorithms. And the effectiveness of the method in the IMUs-based human gait recognition problem is shown, which provides some implications for fine-grained gait recognition.

A robotic grasp detection algorithm based on multiscale features is proposed for autonomous robotic grasping in an unstructured environment. The grasp detection model borrowed the YOLOv3 object detection algorithm and retained the original idea of multiscale detection to improve the perception ability of the grasp rectangle on different scales. Squeeze and excitation blocks were embedded into the Residual Networks (ResNet) structure of the original model, with deformable convolution (DC) introduced, so that the model attained stronger feature extraction ability to cope with more complex grasp detection tasks. Meanwhile, the prediction of the direction angle was transformed into a combination of classification and regression, achieving the prediction of the direction angle of the grabbing frame under different postures. The model was simulated on the Cornell grasp dataset. The results demonstrate that the algorithm in this study can effectively balance the accuracy and efficiency of detection and can migrate the prediction of the grasp rectangle to new objects. The results of online grasp experiments on a Baxter robot show that the average grasp success rate of 93% is achieved for 10 different objects, demonstrating the practical feasibility of the algorithm.