User identification is an essential step in creating a personalised long-term interaction with robots. This requires learning the users continuously and incrementally, possibly starting from a state without any known user. In this article, we describe a multi-modal incremental Bayesian network with online learning, which is the first method that can be applied in such scenarios. Face recognition is used as the primary biometric, and it is combined with ancillary information, such as gender, age, height, and time of interaction to improve the recognition. The Multi-modal Long-term User Recognition Dataset is generated to simulate various human-robot interaction (HRI) scenarios and evaluate our approach in comparison to face recognition, soft biometrics, and a state-of-the-art open world recognition method (Extreme Value Machine). The results show that the proposed methods significantly outperform the baselines, with an increase in the identification rate up to 47.9% in open-set and closed-set scenarios, and a significant decrease in long-term recognition performance loss. The proposed models generalise well to new users, provide stability, improve over time, and decrease the bias of face recognition. The models were applied in HRI studies for user recognition, personalised rehabilitation, and customer-oriented service, which showed that they are suitable for long-term HRI in the real world.

This paper aims to bridge this gap between neuro-symbolic learning (NSL) and graph neural networks (GNN) approaches and provide a comparative study. We argue that the natural evolution of NSL leads to GNNs, while the logic programming foundations of NSL can bring powerful tools to improve the way information is represented and pre-processed for the GNN. In order to make this comparison, we propose HetSAGE, a GNN architecture that can efficiently deal with the resulting heterogeneous graphs that represent typical NSL learning problems. We show that our approach outperforms the state-of-the-art on 3 NSL tasks: CORA, MUTA188 and MovieLens.

We consider the problem of building a state representation model for control, in a continual learning setting. As the environment changes, the aim is to efficiently compress the sensory state information without losing past knowledge, and then use Reinforcement Learning on the resulting features for efficient policy learning. To this end, we propose S-TRIGGER, a general method for Continual State Representation Learning applicable to Variational Auto-Encoders and its many variants. The method is based on Generative Replay, i.e. the use of generated samples to maintain past knowledge. It comes along with a statistically sound method for environment change detection, which self-triggers the Generative Replay. Our experiments on VAEs show that S-TRIGGER learns state representations that allows fast and high-performing Reinforcement Learning, while avoiding catastrophic forgetting. The resulting system has a bounded size and is capable of autonomously learning new information without using past data.

This paper introduces a new raw-audio, environmental, kitchen-related, non-vocal dataset to fill what we consider as a gap in the context of audio datasets. Our so-called Kitchen20 dataset is compared to other datasets such as ESC-50 and shown that both datasets can be merged together in what we call ESC-70. A human quantitative appreciation of the audio samples contained in Kitchen20 is provided as well as several machine learning benchmarks on both Kitchen20 and ESC-70.

Which generative model is the most suitable for Continual Learning? This paper aims at evaluating and comparing generative models on disjoint sequential image generation tasks. We investigate how several models learn and forget, considering various strategies: rehearsal, regularization, generative replay and fine-tuning. We used two quantitative metrics to estimate the generation quality and memory ability. We experiment with sequential tasks on three commonly used benchmarks for Continual Learning (MNIST, Fashion MNIST and CIFAR10). We found that among all models, the original GAN performs best and among Continual Learning strategies, generative replay outperforms all other methods. Even if we found satisfactory combinations on MNIST and Fashion MNIST, training generative models sequentially on CIFAR10 is particularly instable, and remains a challenge. Our code is available online1.

Finding a generally accepted formal definition of a disentangled representation in the context of an agent behaving in an environment is an important challenge towards the construction of data-efficient autonomous agents. Higgins et al. recently proposed Symmetry-Based Disentangled Representation Learning, a definition based on a characterization of symmetries in the environment using group theory. We build on their work and make observations, theoretical and empirical, that lead us to argue that Symmetry-Based Disentangled Representation Learning cannot only be based on static observations: agents should interact with the environment to discover its symmetries. Our experiments can be reproduced in Colab and the code is available on GitHub.

Despite its omnipresence in robotics application, the nature of spatial knowledge and the mechanisms that underlie its emergence in autonomous agents are still poorly understood. Recent theoretical works suggest that the Euclidean structure of space induces invariants in an agent’s raw sensorimotor experience. We hypothesize that capturing these invariants is beneficial for sensorimotor prediction and that, under certain exploratory conditions, a motor representation capturing the structure of the external space should emerge as a byproduct of learning to predict future sensory experiences. We propose a simple sensorimotor predictive scheme, apply it to different agents and types of exploration, and evaluate the pertinence of these hypotheses. We show that a naive agent can capture the topology and metric regularity of its sensor’s position in an egocentric spatial frame without any a priori knowledge, nor extraneous supervision.

Robotic agents, when not equipped with traditional means to capture information about their surroundings, must autonomously learn to extract this information from a very complex environment. In the context of developmental robotics, we use unsupervised representation learning, and more specifically deep autoencoders, in order to capture visual representations. These generic visual representations can in turn be used by different modules of a cognitive architecture. In this paper, we propose a model using these representations to learn an intrinsic measure of bottom-up visual saliency which could later be implemented into a mobile robot.

We present a flexible and scalable architecture that can learn to predict the future behavior of a vehicle in inner-city traffic. While behavior prediction studies have mainly been focusing on lane change events on highways, we apply our approach to a simple inner-city scenario: approaching a traffic light. Our system employs dynamic information about the current ego-vehicle state as well as static information about the scene, in this case position and state of nearby traffic lights.