Structure learning of graphical models for task-oriented robot grasping

In the collective imaginaries a robot is a human like machine as any androids in science fiction. However the type of robots that you will encounter most frequently are machinery that do work that is too dangerous, boring or onerous. Most of the robots in the world are of this type. They can be found in auto, medical, manufacturing and space industries. Therefore a robot is a system that contains sensors, control systems, manipulators, power supplies and software all working together to perform a task. The development and use of such a system is an active area of research and one of the main problems is the development of interaction skills with the surrounding environment, which include the ability to grasp objects. To perform this task the robot needs to sense the environment and acquire the object informations, physical attributes that may influence a grasp. Humans can solve this grasping problem easily due to their past experiences, that is why many researchers are approaching it from a machine learning perspective finding grasp of an object using information of already known objects. But humans can select the best grasp amongst a vast repertoire not only considering the physical attributes of the object to grasp but even to obtain a certain effect. This is why in our case the study in the area of robot manipulation is focused on grasping and integrating symbolic tasks with data gained through sensors. The learning model is based on Bayesian Network to encode the statistical dependencies between the data collected by the sensors and the symbolic task. This data representation has several advantages. It allows to take into account the uncertainty of the real world, allowing to deal with sensor noise, encodes notion of causality and provides an unified network for learning. Since the network is actually implemented and based on the human expert knowledge, it is very interesting to implement an automated method to learn the structure as in the future more tasks and object features can be introduced and a complex network design based only on human expert knowledge can become unreliable. Since structure learning algorithms presents some weaknesses, the goal of this thesis is to analyze real data used in the network modeled by the human expert, implement a feasible structure learning approach and compare the results with the network designed by the expert in order to possibly enhance it.

Centrifuge modeling of earth-reinforced retaining walls

Object of this thesis has been centrifuge modelling of earth reinforced retaining walls with modular blocks facing in order to investigate on the influence of design parameters, such as length and vertical spacing of reinforcement, on the behaviour of the structure. In order to demonstrate, 11 models were tested, each one with different length of reinforcement or spacing. Each model was constructed and then placed in the centrifuge in order to artificially raise gravitational acceleration up to 35 g, reproducing the soil behaviour of a 5 metre high wall. Vertical and horizontal displacements were recorded by means of a special device which enabled tracking of deformations in the structure along its longitudinal cross section, essentially drawing its deformed shape. As expected, results confirmed reinforcement parameters to be the governing factor in the behaviour of earth reinforced structures since increase in length and spacing improved structural stability. However, the influence of the length was found out to be the leading parameter, reducing facial deformations up to five times, and the spacing playing an important role especially in unstable configurations. When failure occurred, failure surface was characterised by the same shape (circular) and depth, regardless of the reinforcement configuration. Furthermore, results confirmed the over-conservatism of codes, since models with reinforcement layers 0.4H long showed almost negligible deformations. Although the experiments performed were consistent and yielded replicable results, further numerical modelling may allow investigation on other issues, such as the influence of the reinforcement stiffness, facing stiffness and varying backfills.

Outdoor bronze conservation: assessment of protective treatments by accelerated aging and of treatment removal procedures by laser cleaning

Outdoor bronzes exposed to the environment form naturally a layer called patina, which may be able to protect the metallic substrate. However, since the last century, with the appearance of acid rains, a strong change in the nature and properties of the copper based patinas occurred [1]. Studies and general observations have established that bronze corrosion patinas created by acid rain are not only disfiguring in terms of loss of detail and homogeneity, but are also unstable [2]. The unstable patina is partially leached away by rainwater. This leaching is represented by green streaking on bronze monuments [3]. Because of the instability of the patina, conservation techniques are usually required. On a bronze object exposed to the outdoor environment, there are different actions of the rainfall and other atmospheric agents as a function of the monument shape. In fact, we recognize sheltered and unsheltered areas as regards exposure to rainwater [4]. As a consequence of these different actions, two main patina types are formed on monuments exposed to the outdoor environment. These patinas have different electrochemical, morphological and compositional characteristics [1]. In the case of sheltered areas, the patina contains mainly copper products, stratified above a layer strongly enriched in insoluble Sn oxides, located at the interface with the uncorroded metal. Moreover, different colors of the patina result from the exposure geometry. The surface color may be pale green for unsheltered areas, and green and mat black for sheltered areas [4]. Thus, in real outdoor bronze monuments, the corrosion behavior is strongly influenced by the exposure geometry. This must be taken into account when designing conservation procedures, since the patina is in most cases the support on which corrosion inhibitors are applied. Presently, for protecting outdoor bronzes against atmospheric corrosion, inhibitors and protective treatments are used. BTA and its derivatives, which are the most common inhibitors used for copper and its alloy, were found to be toxic for the environment and human health [5, 6]. Moreover, it has been demonstrated that BTA is efficient when applied on bare copper but not as efficient when applied on bare bronze [7]. Thus it was necessary to find alternative compounds. Silane-based inhibitors (already successfully tested on copper and other metallic substrates [8]), were taken into consideration as a non-toxic, environmentally friendly alternative to BTA derivatives for bronze protection. The purpose of this thesis was based on the assessment of the efficiency of a selected compound, to protect the bronze against corrosion, which is the 3-mercapto-propyl-trimethoxy-silane (PropS-SH). It was selected thanks to the collaboration with the Corrosion Studies Centre “Aldo Daccò” at the Università di Ferrara. Since previous studies [9, 10, 11] demonstrated that the addition of nanoparticles to silane-based inhibitors leads to an increase of the protective efficiency, we also wanted to evaluate the influence of the addition of CeO2, La2O3, TiO2 nanoparticles on the protective efficiency of 3-mercapto-propyl-trimethoxy-silane, applied on pre-patinated bronze surfaces. This study is the first section of the thesis. Since restorers have to work on patinated bronzes and not on bare metal (except for contemporary art), it is important to be able to recreate the patina, under laboratory conditions, either in sheltered or unsheltered conditions to test the coating and to obtain reliable results. Therefore, at the University of Bologna, different devices have been designed to simulate the real outdoor conditions and to create a patina which is representative of real application conditions of inhibitor or protective treatments. In particular, accelerated ageing devices by wet & dry (simulating the action of stagnant rain in sheltered areas [12]) and by dropping (simulating the leaching action of the rain in unsheltered areas [1]) tests were used. In the present work, we used the dropping test as a method to produce pre-patinated bronze surfaces for the application of a candidate inhibitor as well as for evaluating its protective efficiency on aged bronze (unsheltered areas). In this thesis, gilded bronzes were also studied. When they are exposed to the outside environment, a corrosion phenomenon appears which is due to the electrochemical couple gold/copper where copper is the anode. In the presence of an electrolyte, this phenomenon results in the formation of corrosion products than will cause a blistering of the gold (or a break-up and loss of the film in some cases). Moreover, because of the diffusion of the copper salts to the surface, aggregates and a greenish film will be formed on the surface of the sample [13]. By coating gilded samples with PropS-SH and PropS-SH containing nano-particles and carrying out accelerated ageing by the dropping test, a discussion is possible on the effectiveness of this coating, either with nano-particles or not, against the corrosion process. This part is the section 2 of this thesis. Finally, a discussion about laser treatment aiming at the assessment of reversibility/re-applicability of the PropS-SH coating can be found in section 3 of this thesis. Because the protective layer loses its efficiency with time, it is necessary to find a way of removing the silane layer, before applying a new one on the “bare” patina. One request is to minimize the damages that a laser treatment would create on the patina. Therefore, different laser fluences (energy/surface) were applied on the sample surface during the treatment process in order to find the best range of fluence. In particular, we made a characterization of surfaces before and after removal of PropS-SH (applied on a naturally patinated surface, and subsequently aged by natural exposure) with laser methods. The laser removal treatment was done by the CNR Institute of Applied Physics “Nello Carrara” of Sesto Fiorentino in Florence. In all the three sections of the thesis, a range of non-destructive spectroscopic methods (Scanning Electron Microscopy with Energy Dispersive Spectroscopy (SEM-EDS), μ-Raman spectroscopy, X-Ray diffractometry (XRD)) were used for characterizing the corroded surfaces. AAS (Atomic Absorption Spectroscopy) was used to analyze the ageing solutions from the dropping test in sections 1 and 2.

Deep learning for computer vision: a comparison between convolutional neural networks and hierarchical temporal memories on object recognition tasks

In recent years, Deep Learning techniques have shown to perform well on a large variety of problems both in Computer Vision and Natural Language Processing, reaching and often surpassing the state of the art on many tasks. The rise of deep learning is also revolutionizing the entire field of Machine Learning and Pattern Recognition pushing forward the concepts of automatic feature extraction and unsupervised learning in general. However, despite the strong success both in science and business, deep learning has its own limitations. It is often questioned if such techniques are only some kind of brute-force statistical approaches and if they can only work in the context of High Performance Computing with tons of data. Another important question is whether they are really biologically inspired, as claimed in certain cases, and if they can scale well in terms of "intelligence". The dissertation is focused on trying to answer these key questions in the context of Computer Vision and, in particular, Object Recognition, a task that has been heavily revolutionized by recent advances in the field. Practically speaking, these answers are based on an exhaustive comparison between two, very different, deep learning techniques on the aforementioned task: Convolutional Neural Network (CNN) and Hierarchical Temporal memory (HTM). They stand for two different approaches and points of view within the big hat of deep learning and are the best choices to understand and point out strengths and weaknesses of each of them. CNN is considered one of the most classic and powerful supervised methods used today in machine learning and pattern recognition, especially in object recognition. CNNs are well received and accepted by the scientific community and are already deployed in large corporation like Google and Facebook for solving face recognition and image auto-tagging problems. HTM, on the other hand, is known as a new emerging paradigm and a new meanly-unsupervised method, that is more biologically inspired. It tries to gain more insights from the computational neuroscience community in order to incorporate concepts like time, context and attention during the learning process which are typical of the human brain. In the end, the thesis is supposed to prove that in certain cases, with a lower quantity of data, HTM can outperform CNN.