Department of Computer, Control and Management Engineering

Research StatementIn the future, robots and autonomous devices will become more and morepopular and will finally be part of our everyday life. To achieve thisgoal, robot need to act intelligently and to offer useful services totheir users. Technically, the goal of an intelligent robot is to bringthe environment to a desired configuration by interacting with it.  Toreach this goal the robot might perform complex sequences of actions.To calculate which actions to perform one needs to represent theenvironment at the level of abstraction appropriate for the specifictask.  Our research focuses to develop techniques and models that canbe used in relevant classes of robotic applications, namely:autonomous navigation, object detection and object recognition.  Themethods we developed have been used as building blocks of integratedrobotic systems which can explore the environment or that are used inindustrial applications.  To speed-up the development of robotapplications we also contributed to middle-wares for robotics and weinvestigated how to compare different systems. Learning Models for Autonomous NavigationAcquiring maps with mobile robots has been deeply investigated in thelast two decades.  This problem is known with the name of``Simultaneous Localization And Mapping'' (SLAM).  It consists inestimating simultaneously the map and the position of a mobile robotwhile it moves and measures the environment with its on-board sensors.Within this context, we contributed with approaches that are able toconstruct metric models of the environment by using either filteringor smoothing approaches on graph representations. SLAM with Particle FiltersDuring the first period of our research we investigated particlefilters for SLAM. These filters represent the probability distributionover the possible states of the tracked system via a set of samples.In SLAM, every sample represents a potential robot trajectory.  Weinvestigated how to accurately evolve this distribution of samples inorder to reduce their number, by placing them in the ``right''locations  An opensource implementation of these algorithms (GMapping ) became well known to therobotic community. To improve the efficiency of these algorithms, we subsequentlyinvestigated on more compact representations that can exploit thesymmetries in the environment and in the SLAMprocess.  These systems have beessuccessfully used in search and rescueapplications onhumanoids. They also have been extended tofixed-lag filters or to operate onmixed representations. Graph-based SLAMBased on the experience with particle filter SLAM, second phase of ourresearch we developed graph-based SLAM algorithms.  As the namesuggests, these approaches model the problem as a graph. Each node ofthe graph represents a local map, and the position of this local mapwithin a global reference system. An edge between two nodes representsa spatial constraint between two local maps and is labeled with thespatial transformation obtained by aligning the two local maps.  Thus,solving graph-based SLAM means:

• constructing the graph and the labels from the raw sensor data (graph construction)  and
• find the locations of the local maps that better satisfies the  constraints encoded in the edges (graph optimization).

To construct the graph we developed efficient algorithms to alignlocal maps made of 2D laser scans, on images andinertial data or on 3Dscans. To reject wrong alignments we developedalgorithms to efficiently compute the relative uncertainty between twonodes.  To optimize the graph we developed an algorithm based onstochastic gradient descent.This algorithm relies on a tree parametrization of the graph to speedup the convergence.  An open-source implementation has been madeavailable TORO ) and it isactually used by many research groups, for instance Willow Garage,Carnegie Mellon University, University of T'uebingen, University of Orebro,University of Parma and others. When the robot moves continuously in the environment the full SLAMproblem in its naive formulation becomes not feasible due to the largeamount of data that needs to be processed at each cycle. To this endwe investigated the possibility of dropping the less informativenodes. Furthermore, we approached theoptimization of these large scale problems by using hierarchicalrepresentations HOGMAN). Object Detection and Object ModelingTo perform complex tasks in unstructured environments, a robot shouldbe able to identify the known objects that it needs to manipulate.  Inparticular, to manipulate an object a robot needs to find theinstances and the locations of this object in a scene starting from aknown model.  In a recent work, we approached the problem of detectingknown objects in a 3D scene.  The algorithmrelies on a compact representation of the three dimensional data(range images), and utilizes a variant of RANSAC to align a knownmodel on a portion of the scene. Subsequently, we developed analgorithm that learns models from the scenes in an unsupervisedway. This is done by aligning a portion of the scene onto another.The recurrent consistent portions are combined to form models.  Thisproblem has substantial overlap with multi-robot SLAM Autonomous ExplorationMoving autonomously is an essential skill for any application thatinvolves mobile robots. To this end one needs a map of theenvironment.  Thus, having a device which is able to autonomouslygather such an information is seen as highly relevant by the roboticcommunity. This task is called autonomous exploration.Traditional SLAM algorithms are substantially passive, in the sensethat they operate on a stream of data without controlling how thesedata are taken.  However, it is known that the particular trajectoryfollowed by the robot has a substantial effect on the outcome of SLAM.We introduced an entropy based measure to drive the robot.  This systemtakes into account the effects that a particular observation has onthe outcome of SLAM and chooses the one which maximizes theexpected information gain  about the environment. Integration and Comparison of SubsystemsAn important aspect of our research is the systematic comparison andthe evaluation of different approaches in different scenarios.We proposed a metric to compare SLAMalgorithms which relates the quality of the map to the energy requiredto ``deform" it to obtain the ground truth.  To quickly builddifferent applications involving mobile robots, we contributed to opensource systems with the realization of different components.  Wefurthermore contributed to the design and implementation of anavigation system for small-size helicopters available at (http://www.openquadrotor.org).