In this paper, we propose an active learning approach applied to a music performance imitation scenario. The humanoid robot iCub listens to a human performance and then incrementally learns to use a virtual musical instrument in order to imitate the given sequence. This is achieved by first learning a model of the instrument, needed to locate where the required sounds are heard in a virtual keyboard layed out in a tactile interface. Then, a model of its body capabilities is also learnt, which serves to establish the likelihood of success of the actions needed to imitate the sequence of sounds and to correct the errors made by the underlying kinematic controller. It also uses self-evaluation stages to provide feedback to the human instructor, which can be used to guide its learning process.

Nowadays robots are becoming more ubiquitous, and focus is increasingly put on the variety of tasks they can perform autonomously. However, due to the dynamics of the environment or the robot itself, sometimes the models that were learned in the past do not exactly fit in the present situation. Adaptation mechanisms are of key importance in this case since they enable the robot to reuse prior knowledge, which is useful for the current context. In this work we introduce an active adaptation mechanism which enables a humanoid robot to recover from a failure, exemplified as a displacement of the object it is interacting with.

Gaussian Mixture Models have been widely used in robotic control and in sensory anticipation applications. A mixture model is learnt from demonstrations and later used to infer the most likely control signals, or is also used as a forward model to predict the change in sensory signals over time. However, such models often are too big to be tractable in real-time applications. In this paper we introduce the Context-GMM, a method to learn sparse priors over the mixture components. Such priors are stable over large amounts of time and provide a way of selecting very small subsets of mixture components without significant loss in accuracy and with huge computational savings.

In order to anticipate dangerous events, like a collision, an agent needs to make long-term predictions. However, those are challenging due to uncertainties in internal and external variables and environment dynamics. A sensorimotor model is acquired online by the mobile robot using a state-of-the-art method that learns the optical flow distribution in images, both in space and time. The learnt model is used to anticipate the optical flow up to a given time horizon and to predict an imminent collision by using reinforcement learning. We demonstrate that multi-modal predictions reduce to simpler distributions once actions are taken into account.

In this paper we study the mechanisms that enable an active agent to make long-term predictions of optical flow with a model that is learned dynamically. We analyse the optical flow distribution in terms of space and time, that is, what are the experienced optical flow values and how do they change in time. We show how complex the posterior distributions become when long-term predictions are needed, which breaks time-consistency assumption. The choice of one predictor or another should be made in terms of how the data is distributed. Moreover, we use a generic state-of-the-art incremental online learning algorithm [8] for the task of building a model to predict the optical flow perceived by a mobile robot. Finally, as an application, the model is also used to learn a simple predictor for anticipating an imminent collision.

Approaches to object localization based on codebooks do not exploit the dependencies between appearance and geometric information present in training data. This work addresses the problem of computing a codebook tailored to the task of localization by applying regularization based on geometric information. We present a novel method, the Regularized Combined Partitional-Agglomerative clustering, which extends the standard CPA method by adding extra knowledge to the clustering process to preserve as much geometric information as needed. Due to the time complexity of the methodology, we also present an implementation on the GPU using nVIDIA CUDA technology, speeding up the process with a factor

Object recognition has been widely researched for several decades and in the recent years new methods capable of general object classification have appeared. However very few work has been done to adapt these methods to the challenges raised by mobile robotics. In this article we discuss the data sources (appearence information, temporal context, etc.) that such methods could use and we review several state of the art object recognition methods that build in one or more of these sources. Finally we run an object based robot localization experiment using an state of the art object recognition method and we show that good results are obtained even with a naive place descriptor.