Physical Interaction of Autonomous Robots in Complex Environments

Recent breakthroughs in the fields of computer vision and robotics are firmly changing the people perception about robots. The idea of robots that substitute humansisnowturningintorobotsthatcollaboratewiththem. Serviceroboticsconsidersrobotsaspersonalassistants. Itsafelyplacesrobotsindomesticenvironments in order to facilitate humans daily life. Industrial robotics is now reconsidering its basic idea of robot as a worker. Currently, the primary method to guarantee the personnels safety in industrial environments is the installation of physical barriers around the working area of robots. The development of new technologies and new algorithms in the sensor field and in the robotic one has led to a new generation of lightweight and collaborative robots. Therefore, industrial robotics leveraged the intrinsic properties of this kind of robots to generate a robot co-worker that is able to safely coexist, collaborate and interact inside its workspace with both personnels and objects. This Ph.D. dissertation focuses on the generation of a pipeline for fast object pose estimation and distance computation of moving objects,in both structured and unstructured environments,using RGB-D images. This pipeline outputs the command actions which let the robot complete its main task and fulfil the safety human-robot coexistence behaviour at once. The proposed pipeline is divided into an object segmentation part,a 6D.o.F. object pose estimation part and a real-time collision avoidance part for safe human-robot coexistence. Firstly, the segmentation module finds candidate object clusters out of RGB-D images of clutter scenes using a graph-based image segmentation technique. This segmentation technique generates a cluster of pixels for each object found in the image. The candidate object clusters are then fed as input to the 6 D.o.F. object pose estimation module. The latter is in charge of estimating both the translation and the orientation in 3D space of each candidate object clusters. The object pose is then employed by the robotic arm to compute a suitable grasping policy. The last module generates a force vector field of the environment surrounding the robot, the objects and the humans. This force vector field drives the robot toward its goal while any potential collision against objects and/or humans is safely avoided. This work has been carried out at Politecnico di Torino, in collaboration with Telecom Italia S.p.A.