Robotic obstacle avoidance in cluttered and dense environments is an important issue in robotic navigation. Over the past few years a number of techniques has been proposed to deal with safe navigation among obstacles in unknown scenarios. Unfortunately many of these methods do not consider obstacle velocities, which can rise some serious questions concerning their safety. This paper will deal with a novel approach to moving obstacle avoidance in holonomic robots. It proposes the Forbidden Velocity Map, a generalization of the Dynamic Window concept that considers obstacle and robot shape, velocity and dynamics, resulting in a safe, reactive real-time navigation algorithm that is able to deal with navigation in unpredictable and cluttered scenarios.

This work proposes a new system capable of real-time ship instance segmentation during maritime surveillance missions by unmanned aerial vehicles using an onboard standard RGB camera. The implementation requires two stages: an instance segmentation network able to produce fast and reliable preliminary segmentation results and a post-processing 3D fully connected Conditional Random Field, which significantly improves segmentation results by exploring temporal correlations between nearby frames in video sequences. Moreover, due to the absence of maritime datasets consisting of properly labeled video sequences, we create a new dataset comprising synthetic video sequences of maritime surveillance scenarios (MarSyn). The main advantages of this approach are the possibility of generating a vast set of images and videos, being able to represent real-world scenarios without the necessity of deploying the real vehicle, and automatic labels, which eliminate human labeling errors. We train the system with the MarSyn dataset and with aerial footage from publicly available annotated maritime datasets to validate the proposed approach. We present some experimental results and compare them to other approaches, and we also illustrate the temporal stability provided by the second stage in missing frames and wrong segmentation scenarios.

Autonomous service robots have become a key research topic in robotics, particularly for household chores. A typical home scenario is highly unconstrained and a service robot needs to adapt constantly to new situations. In this paper, we address the problem of autonomous cleaning tasks in uncontrolled environments. In our approach, a human instructor uses kinestethic demonstrations to teach a robot how to perform different cleaning tasks on a table. Then, we use Task Parametrized Gaussian Mixture Models (TP-GMMs) to encode the demonstrations variability, while providing appropriate generalization abilities. TP-GMMs extend Gaussian Mixture Models with an auxiliary set of reference frames, in order to extrapolate the demonstrations to different task parameters such as movement locations, amplitude or orientations. However, the reference frames (that parametrize TP-GMMs) can be very difficult to extract in practice, as it may require segmenting the cluttered images of the working table-top. Instead, in this work the reference frames are automatically extracted from robot camera images, using a deep neural network that was trained during human demonstrations of a cleaning task. This approach has two main benefits: (i) it takes the human completely out of the loop while performing complex cleaning tasks; and (ii) the network is able to identify the specific task to be performed directly from image data, thus also enabling automatic task selection from a set of previously demonstrated tasks. The system was implemented on the iCub humanoid robot. During the tests, the robot was able to successfully clean a table with two different types of dirt (wiping a marker's scribble or sweeping clusters of lentils).

This paper proposes a supervised algorithm for online learning of input-output relations that is particularly suitable to simultaneously learn the forward and inverse kinematics of general manipulators - the multi-valued nature of the inverse kinematics of serial chains and forward kinematics of parallel manipulators makes it infeasible to apply state-of-the-art learning techniques to these problems, as they typically assume a single-valued function to be learned. The proposed algorithm is based on a generalized expectation maximization approach to fit an infinite mixture of linear experts to an online stream of data samples, together with an outlier probabilistic model that dynamically grows the number of linear experts allocated to the mixture, this way controlling the complexity of the resulting model. The result is an incremental, online and localized learning algorithm that performs nonlinear, multivariate regression on multivariate outputs by approximating the target function by a linear relation within each expert input domain, which can directly provide forward and inverse multi-valued estimates. The experiments presented in this paper show that it can achieve, for single-valued functions, a performance directly comparable to state-of-the-art online function approximation algorithms, while additionally providing inverse predictions and the capability to learn multi-valued functions in a natural manner. To our knowledge this is a distinctive property of the algorithm presented in this paper.

Accurate dynamic models can be very difficult to compute analytically for complex robots; moreover, using a precomputed fixed model does not allow to cope with unexpected changes in the system. An interesting alternative solution is to learn such models from data, and keep them up-to-date through online adaptation. In this paper we consider the problem of learning the robot inverse dynamic model under dynamically varying contexts: the robot learns incrementally and autonomously the model under different conditions, represented by the manipulation of objects of different weights, that change the dynamics of the system. The inverse dynamic mapping is modeled as a multi-valued function, in which different outputs for the same input query are related to different dynamic contexts (i.e. different manipulated objects). The mapping is estimated using IMLE, a recent online learning algorithm for multi-valued regression, and used for Computed Torque control. No information is given about the context switch during either learning or control, nor any assumption is made about the kind of variation in the dynamics imposed by a new contexts. Experimental results with the iCub humanoid robot are provided.

We address the problem of teaching a robot how to autonomously perform table-cleaning tasks in a robust way. In particular, we focus on wiping and sweeping a table with a tool (e.g., a sponge). For the training phase, we use a set of kinesthetic demonstrations performed over a table. The recorded 2D table-space trajectories, together with the images acquired by the robot, are used to train a deep convolutional network that automatically learns the parameters of a Gaussian Mixture Model that represents the hand movement. After the learning stage, the network is fed with the current image showing the location/shape of the dirt or stain to clean. The robot is able to perform cleaning arm-movements, obtained through Gaussian Mixture Regression using the mixture parameters provided by the network. Invariance to the robot posture is achieved by applying a plane-projective transformation before inputting the images to the neural network; robustness to illumination changes and other disturbances is increased by considering an augmented data set. This improves the generalization properties of the neural network, enabling for instance its use with the left arm after being trained using trajectories acquired with the right arm. The system was tested on the iCub robot generating a cleaning behaviour similar to the one of human demonstrators.