Deep Learning (DL) based methods for object detection achieve remarkable performance at the cost of computationally expensive training and extensive data labeling. Robots embodiment can be exploited to mitigate this burden by acquiring automatically annotated training data via a natural interaction with a human showing the object of interest, hand-held. However, learning solely from this data may introduce biases (the so-called domain shift), and prevents adaptation to novel tasks. While Weakly-supervised Learning offers a well-established set of techniques to cope with these problems in general-purpose Computer Vision, its adoption in challenging robotic domains is still at a preliminary stage. In this work, we target the scenario of a robot trained in a teacher-learner setting to detect handheld objects. The aim is to improve detection performance in different settings by letting the robot explore the environment with a limited human labeling budget. We compare several techniques for WSL in detection pipelines to reduce model re-training costs without compromising accuracy, proposing solutions which target the considered robotic scenario. We show that the robot can improve adaptation to novel domains, either by interacting with a human teacher (Active Learning) or with an autonomous supervision (Semi-supervised Learning). We integrate our strategies into an on-line detection method, achieving efficient model update capabilities with few labels. We experimentally benchmark our method on challenging robotic object detection tasks under domain shift1.

In this paper, we address the problem of enhancing the speech of a speaker of interest in a cocktail party scenario when visual information of the speaker of interest is available.Contrary to most previous studies, we do not learn visual features on the typically small audio-visual datasets, but use an already available face landmark detector (trained on a separate image dataset).The landmarks are used by LSTM-based models to generate time-frequency masks which are applied to the acoustic mixed-speech spectrogram. Results show that: (i) land-mark motion features are very effective features for this task, (ii) similarly to previous work, reconstruction of the target speaker's spectrogram mediated by masking is significantly more accurate than direct spectrogram reconstruction, and (iii) the best masks depend on both motion landmark features and the input mixed-speech spectrogram.To the best of our knowledge, our proposed models are the first models trained and evaluated on the limited size GRID and TCD-TIMIT datasets, that achieve speaker-independent speech enhancement in a multi-talker setting.

Autonomous robots can rely on attention mechanisms to explore complex scenes and select salient stimuli relevant for behaviour. Stimulus selection should be fast to efficiently allocate available (and limited) computational resources to process in detail a subset of the otherwise overwhelmingly large sensory input. The amount of processing required is a product of the amount of data sampled by a robot's sensors; while a standard RGB camera produces a fixed amount of data for every pixel of the sensor, an event-camera produces data only for where there is a contrast change in the field of view, and does so with a lower latency. In this paper, we describe the implementation of a state-of-the-art bottom-up attention model, based on structuring the visual scene in terms of proto-objects. As an event-camera encodes different visual information compared to frame-based cameras, the original algorithm must be adapted and modified. We find that the event-camera's inherent detection of edges removes the need for some early stages of processing in the model. We describe the modifications, compare the event-driven algorithm to the original, and validate the potential for use on the iCub humanoid robot.

To precisely reach for an object with a humanoid robot, it is of central importance to have good knowledge of both end-effector, object pose and shape. In this work we propose a framework for markerless visual servoing on unknown objects, which is divided in four main parts: I) a leastsquares minimization problem is formulated to find the volume of the object graspable by the robot's hand using its stereo vision; II) a recursive Bayesian filtering technique, based on Sequential Monte Carlo (SMC) filtering, estimates the 6D pose (position and orientation) of the robot's end-effector without the use of markers; III) a nonlinear constrained optimization problem is formulated to compute the desired graspable pose about the object; IV) an image-based visual servo control commands the robot's end-effector toward the desired pose. We demonstrate effectiveness and robustness of our approach with extensive experiments on the iCub humanoid robot platform, achieving real-time computation, smooth trajectories and subpixel precisions.

Future humanoid robots will be expected to carry out a wide range of tasks for which they had not been originally equipped by learning new skills and adapting to their environment. A crucial requirement towards that goal is to be able to take advantage of external elements as tools to perform tasks for which their own manipulators are insufficient; the ability to autonomously learn how to use tools will render robots far more versatile and simpler to design. Motivated by this prospect, this paper proposes and evaluates an approach to allow robots to learn tool affordances based on their 3D geometry. To this end, we apply tool-pose descriptors to represent tools combined with the way in which they are grasped, and affordance vectors to represent the effect tool-poses achieve in function of the action performed. This way, tool affordance learning consists in determining the mapping between these 2 representations, which is achieved in 2 steps. First, the dimensionality of both representations is reduced by unsupervisedly mapping them onto respective Self-Organizing Maps (SOMs). Then, the mapping between the neurons in the tool-pose SOM and the neurons in the affordance SOM for pairs of tool-poses and their corresponding affordance vectors, respectively, is learned with a neural based regression model. This method enables the robot to accurately predict the effect of its actions using tools, and thus to select the best action for a given goal, even with tools not seen on the learning phase.

This paper addresses recursive markerless estimation of a robot's end-effector using visual observations from its cameras. The problem is formulated into the Bayesian framework and addressed using Sequential Monte Carlo (SMC) filtering. We use a 3D rendering engine and Computer Aided Design (CAD) schematics of the robot to virtually create images from the robot's camera viewpoints. These images are then used to extract information and estimate the pose of the end-effector. To this aim, we developed a particle filter for estimating the position and orientation of the robot's end-effector using the Histogram of Oriented Gradient (HOG) descriptors to capture robust characteristic features of shapes in both cameras and rendered images. We implemented the algorithm on the iCub humanoid robot and employed it in a closed-loop reaching scenario. We demonstrate that the tracking is robust to clutter, allows compensating for errors in the robot kinematics and servoing the arm in closed loop using vision.

In recent years the robotics field has witnessed an interesting new trend. Several companies started the production of service robots whose aim is to cooperate with humans. The robots developed so far are either rather expensive or unsuitable for manipulation tasks. This article presents the result of a project which wishes to demonstrate the feasibility of an affordable humanoid robot. R1 is able to navigate, and interact with the environment (grasping and carrying objects, operating switches, opening doors etc). The robot is also equipped with a speaker, microphones and it mounts a display in the head to support interaction using natural channels like speech or (simulated) eye movements. The final cost of the robot is expected to range around that of a family car, possibly, when produced in large quantities, even significantly lower. This goal was tackled along three synergistic directions: use of polymeric materials, light-weight design and implementation of novel actuation solutions. These lines, as well as the robot with its main features, are described hereafter.

The ability to learn about and efficiently use tools constitutes a desirable property for general purpose humanoid robots, as it allows them to extend their capabilities beyond the limitations of their own body. Yet, it is a topic that has only recently been tackled from the robotics community. Most of the studies published so far make use of tool representations that allow their models to generalize the knowledge among similar tools in a very limited way. Moreover, most studies assume that the tool is always grasped in its common or canonical grasp position, thus not considering the influence of the grasp configuration in the outcome of the actions performed with them. In the current paper we present a method that tackles both issues simultaneously by using an extended set of functional features and a novel representation of the effect of the tool use. Together, they implicitly account for the grasping configuration and allow the iCub to generalize among tools based on their geometry. Moreover, learning happens in a self-supervised manner: First, the robot autonomously discovers the affordance categories of the tools by clustering the effect of their usage. These categories are subsequently used as a teaching signal to associate visually obtained functional features to the expected tool's affordance. In the experiments, we show how this technique can be effectively used to select, given a tool, the best action to achieve a desired effect.

We present a novel method for three-finger precision grasp and its implementation in a complete grasping tool-chain. We start from binocular vision to recover the partial 3D structure of unknown objects. We then process the incomplete 3D point clouds searching for good triplets according to a function that accounts for both the feasibility and the stability of the solution. In particular, while stability is determined using the classical force-closure approach, feasibility is evaluated according to a new measure that includes information about the possible configuration shapes of the hand as well as the hand's inverse kinematics. We finally extensively assess the proposed method using the stereo vision and the kinematics of the iCub robot.

Systematically developing high-quality reusable software components is a difficult task and requires careful design to find a proper balance between potential reuse, functionalities and ease of implementation. Extendibility is an important property for software which helps to reduce cost of development and significantly boosts its reusability. This work introduces an approach to enhance components reusability by extending their functionalities using plug-ins at the level of the connection points (ports). Application-dependent functionalities such as data monitoring and arbitration can be implemented using a conventional scripting language and plugged into the ports of components. The main advantage of our approach is that it avoids to introduce application-dependent modifications to existing components, thus reducing development time and fostering the development of simpler and therefore more reusable components. Another advantage of our approach is that it reduces communication and deployment overheads as extra functionalities can be added without introducing additional modules. The details of the plug-in system is described in the paper and its advantages for the development of robotics applications are demonstrated by developing a step-by-step example on the iCub humanoid robot.

For any robot, the ability to recognize and manipulate unknown objects is crucial to successfully work in natural environments. Object recognition and categorization is a very challenging problem, as 3-D objects often give rise to ambiguous, 2-D views. Here, we present a perception-driven exploration and recognition scheme for in-hand object recognition implemented on the iCub humanoid robot. In this setup, the robot actively seeks out object views to optimize the exploration sequence. This is achieved by regarding the object recognition problem as a localization problem. We search for the most likely viewpoint position on the viewsphere of all objects. This problem can be solved efficiently using a particle filter that fuses visual cues with associated motor actions. Based on the state of the filter, we can predict the next best viewpoint after each recognition step by searching for the action that leads to the highest expected information gain. We conduct extensive evaluations of the proposed system in simulation as well as on the actual robot and show the benefit of perception-driven exploration over passive, vision-only processes at discriminating between highly similar objects. We demonstrate that objects are recognized faster and at the same time with a higher accuracy.

Interaction with its environment is a key requisite for a humanoid robot. Especially the ability to recognize and manipulate unknown objects is crucial to successfully work in natural environments. Visual object recognition, however, still remains a challenging problem, as three-dimensional objects often give rise to ambiguous, two-dimensional views. Here, we propose a perception-driven, multisensory exploration and recognition scheme to actively resolve ambiguities that emerge at certain viewpoints. We define an efficient method to acquire two-dimensional views in an object-centered task space and sample characteristic views on a view sphere. Information is accumulated during the recognition process and used to select actions expected to be most beneficial in discriminating similar objects. Besides visual information we take into account proprioceptive information to create more reliable hypotheses. Simulation and real-world results clearly demonstrate the efficiency of active, multisensory exploration over passive, visiononly recognition methods.

In this paper we discuss the use of the infinite Gaussian mixture model and Dirichlet processes for learning robot movements from demonstrations. Starting point of this work is an earlier paper where the authors learn a non-linear dynamic robot movement model from a small number of observations. The model in that work is learned using a classical finite Gaussian mixture model (FGMM) where the Gaussian mixtures are appropriately constrained. The problem with this approach is that one needs to make a good guess for how many mixtures the FGMM should use. In this work, we generalize this approach to use an infinite Gaussian mixture model (IGMM) which does not have this limitation. Instead, the IGMM automatically finds the number of mixtures that are necessary to reflect the data complexity. For use in the context of a non-linear dynamic model, we develop a Constrained IGMM (CIGMM). We validate our algorithm on the same data that was used in [5], where the authors use motion capture devices to record the demonstrations. As further validation we test our approach on novel data acquired on our iCub in a different demonstration scenario in which the robot is physically driven by the human demonstrator.

Cross-situational learning is based on the idea that a learner can determine the meaning of a word by finding something in common across all observed uses of that word. Although cross-situational learning is usually modeled through stochastic guessing games in which the input data vary erratically with time (or rounds of the game), here we investigate the possibility of applying the deterministic neural modeling fields (NMF) categorization mechanism to infer the correct object-word mapping. Two different representations of the input data were considered. The first is termed object-word representation because it takes as inputs all possible object-word pairs and weighs them by their frequencies of occurrence in the stochastic guessing game. A re-interpretation of the problem within the perspective of learning with noise indicates that the cross-situational scenario produces a too low signal-to-noise ratio, explaining thus the failure of NMF to infer the correct object-word mapping. The second representation, termed context-word, takes as inputs all the objects that are in the pupil's visual field (context) when a word is uttered by the teacher. In this case we show that use of two levels of hierarchy of NMF allows the inference of the correct object-word mapping.

Neural networks have been proposed as an ideal cognitive modeling methodology to deal with the symbol grounding problem. More recently, such neural network approaches have been incorporated in studies based on cognitive agents and robots. In this paper we present a new model of symbol grounding transfer in cognitive robots. Language learning simulations demonstrate that robots are able to acquire new action concepts via linguistic instructions. This is achieved by autonomously transferring the grounding from directly grounded action names to new higher-order composite actions. The robot's neural network controller permits such a grounding transfer. The implications for such a modeling approach in cognitive science and autonomous robotics are discussed.