Search

Service robots and other smart devices, such as smartphones, have access to large amounts of data suitable for learning models, which can greatly improve the customer experience. Federated learning is a popular framework that allows multiple distributed devices to train deep learning models remotely, collaboratively, and preserving data privacy. However, little research has been done regarding the scenario where data distribution is non-identical among the participants and it also changes over time in unforeseen ways, causing what is known as concept drift. This situation is, however, very common in real life, and poses new challenges to both federated and continual learning. In this work, we propose an extension of the most widely known federated algorithm, FedAvg, adapting it for continual learning under concept drift. We empirically demonstrate the weaknesses of regular FedAvg and prove that our extended method outperforms the original one in this type of scenario.

Service robots at homes or works are expected to upload data that can be used by companies to fix the controllers and improve robot behaviours. Nevertheless, this is a delicate issue that concerns data privacy. Instead, we propose an iterative process of local learning (in the robots) and global consensus (in the cloud) that still preserves the benefits of learning from the crowd but when models instead of data are uploaded to a server. This strategy is also valid for mobile phones or other devices. In fact, in order to work with a heterogeneous community of users, we have applied our strategy in a real problem with mobile phones: walking recognition. We achieved very high performances without the need of massive amounts of centralized data.

Nowadays, mobile phones are used more and more for purposes that have nothing to do with phone calls or simple data transfers. One example is indoor inertial navigation. Within this task, a central problem is to obtain a good estimation of the user heading, robust to magnetic interference and changes in the position of the mobile device with respect to the user. In this paper we propose a method able to provide a robust user heading as a result of detecting the relative position of the mobile phone with respect to the user, together with a heuristic computation of the heading from different Euler representations. We have performed an experimental validation of our proposal comparing it with the Android default compass. The results confirm the good performance of our method.

This paper describes the challenges that involve developing a software capable of capturing users’ faces on mobile devices in a non-collaborative environment. The goal is to generate a set of quality training samples of the user’s face for the construction of a model that can be used in a later phase of biometric identification. To this end, a supervised learning system is integrated to determine when a photo should be taken. This learning is supported by a varied input data set that contains information regarding the pose of the device, its manipulation and other environmental factors such as lighting. The software also has different ways of working with the objective of not wasting resources and be little invasive. Working modes are managed with an easy-to-maintain and scalable rules-based system. The experimental results show the robustness of the proposal.

Navigation is an essential feature for smartphones, even indoors. Having a robust step count algorithm is the cornerstone for building an inertial navigator based on accelerometer sensors. However, accelerometer data is very sensitive to body movements, so separating noise from real steps is not a trivial issue. Our main hypothesis is that Mean Squared Error (MSE) measured between predicted and real signal gives a clear distinction between ideal steps, noisy steps and pure noise. In this paper we propose a combination of techniques to obtain a robust step count model for smartphones. Using the vertical component of the acceleration, Support Vector Regression (SVR) for modeling user’s activity and an algorithm that combines peak-valley detection with high MSE steps filtering, we achieve a computational efficient and robust model for detecting steps.

In this paper we describe an strategy to obtain a robust pedometer in mobile phones through a learning process that is carried out in the mobile itself. Using the vertical component of the acceleration, dynamic time warping and data collected on the mobile, we achieve a model able to detect steps and which exhibits an important robustness to the way the mobile is being carried out. We believe this robustness is due to the fact that the model, learnt on the mobile, requires less heuristic parameters and is linked to specific characteristics of the user and the hardware. We have tested our strategy in real experiments carried out at our research centre.

This paper describes the challenges that involve playing with the Nao humanoid robot on a tablet. For that purpose, an inverse kinematic solver that allows the robot to move it’s limbs, and a computer vision algorithm that allows the robot to understand the items displayed on the tablet, are needed. The presented solution uses NAOqi’s Cartesian Control and OpenCV’s Hough Transform respectively. To overcome the lack of force and tactile sensors on Nao’s hand, we propose a touch movement based on visual feedback. As an initial approach, we chose the Tic-Tac-Toe game and we introduced interaction mechanisms to make it more pleasant and enjoyable, with the objective of creating a template for HRI and machine learning integration. The experimental results show the robustness of the proposed architecture.

The humanoid robot Nao is a great platform for robotics research, in particular it provides an important testbed for computer vision, machine learning and human robot interface. Nevertheless, its limited sensorization and computation power reduces its autonomy severely. To overcome some of these limitations, in this paper we describe the integration of a RGB-D camera together with a mini-PC, into the Nao robot. Our objective is to get a Nao robot being able to carry out autonomously (onboard) all the tasks that involve 3D environment perception. As an example we used two applications: (1) mimic of human movements, (which will be used on learning by demonstration), and (2) RTABmap SLAM algorithm. Finally, we also tested the Nao’s walking stability when it was equipped with all the new elements.

Mobile robotics has achieved important progress and level of maturity. Nevertheless, to increase the complexity of the tasks that mobile robots can perform in indoor environments, we need to provide them with a scene understanding of their surrounding. Scene recognition usually involves building image classifiers using training data. These classifiers work with features extracted from the images to recognize different categories. Later on, these classifiers can be used to label any image taken by the robot. The problem is that the training data used to recognize the scene might be redundant and noisy, thus reducing significantly the performance of the classifiers. To avoid this, we propose an unsupervised algorithm able to recognize when an image is unrepresentative, redundant or outlier. We have tested our algorithm in real and difficult environments achieving very promising results which take us a step closer to a complete unsupervised scene recognition with high accuracy.

Future robot systems will perform increasingly complex tasks in decreasingly well-structured and known environments. Robots will need to adapt their hardware and software, first only to foreseen, but ultimately to more complex changes of the environment. In this paper we describe a learning strategy based on reinforcement which allows fast robot learning from scratch using only its interaction with the environment, even when the reward is provided by a human observer and therefore is highly non-deterministic and noisy. To get this our proposal uses a novel representation of the action space together with an ensemble of learners able to forecast the time interval before a robot failure

Scene understanding is still an important challenge in robotics. In this paper we analyze the utility of scene recognition to determine the localization of a robot. We assume that multi-sensor localization systems may be very useful in crowded environments where there will be many people around the robot but not many changes of the furniture. In our localization system we categorize the sensors in two groups: accurate sensor models able to determine the pose of the robot accurately but which are sensible to noise or the presence of people. Robust sensor modalities able to provide rough information about the pose of the robot in almost any condition. The performance of our localization strategy was analyzed through two experiments realized in the Centro Singular de Investigacion en Tecnoloxias da Informacion (CITIUS), at the University of Santiago de Compostela.

Scene understanding is still an important challenge in robotics. In this paper we analyse the impact of several global and local image representations to solve the task of scene recognition. The performance of the different alternatives were compared using a two benchmarks of images: (a) the public database KTH_IDOL and, (b) a base of images taken in the Centro Singular de Investigacion en Tecnoloxias da Informacion (CITIUS), at the University of Santiago de Compostela. The results are promising not only regarding the accuracy achieved, but mostly because we have found a combination of an holistic representation and local information that allows a correct classification of images robust to specular reflections, illumination conditions, changes of viewpoint, etc.

Scene understanding is still an important challenge in robotics. Nevertheless scene recognition involves determining when an image is good enough to represent the scene and therefore it can be used for classification. Most research on scene recognition involves working with sets of images which have been acquired using a predefined sampling rate, nevertheless, this means working with very noisy and redundant sets of images. In this paper we analyse different alternatives to automatically select images according to amount of information they provide and how representative they are.