Past Research - 2009 - 2017

During my experience in both NIFTI and TRADR projects, I had the chance to deeply understand the main problems to be addressed, when robots are involved, together with humans, in rescue missions. Closely working with fire-fighters, I better understood what the capabilities of a robot should be, in terms of autonomy, in order to support rescue responders. According to this experience, we worked on increasing the level of autonomy of a rescue robot. We also had two in-field experiences, where we developed a robotic system to assess damage to historical buildings, and cultural artifacts located therein, first in July 2012, after the earthquake in Mirandola, in the Emilia-Romagna region, Northern Italy, and second in September 2016, after the earthquake in Amatrice, Lazio Region, Central Italy.

We developed a framework for real-time 3D motion planning and control of tracked robots, for autonomous navigation in harsh environments. This framework is based on a semantic representation of the environment. This representation is used to plan feasible 3D paths for the robot, toward a target position. The physical execution of the path is delegated to a decoupled controller, responsible of both generating velocity commands and adapting the robot morphology during the tracking task. We extended the controller with a statistical model assessing the touch of the articulated mechanical components of the robot with the traversed surfaces. This model is used to both correct the estimation of the robot morphology and to ensure that the robot has a better traction onto harsh terrain. We also worked on improving the semantic representation of the environment with information about traversability. We finally improved the path planning algorithm to also take into account dynamic obstacles.

We developed a robotic system that employs high-level control in order to operate in a real-world setting where the main task is human-assisted exploration of an environment. In this system, we integrated multi-modal perception from vision and mapping with a model-based executive control. In this framework, action planning is performed using a high-level representation of the environment that is obtained through topological segmentation of the metric map and object detection and 3D localisation in the map. This work contributes in providing an effective method to build a logical representation of the low-level perception of the robot and, compile the robot perception into knowledge and, finally, integrate this knowledge with a model-based executive control. The overall system enables the robot to infer strategies in so generating parametric plans that are directly instantiated from perception.

We also worked on a new approach for robot cognitive control based on modelling robot stimuli, the stimulus-response and the resulting task switching or stimulus inhibition. The proposed framework contributes to the state of the art on robot planning and high level control as it provides a novel view on robot behaviours and a complete implementation on an advanced robotic platform.

This research concerns the training and testing of a Convolutional Neural Network (CNN) to assess the traversability of the terrain for safe autonomous navigation of an articulated tracked robot in urban search and rescue domain applications. Traversability assessment is fundamental under different fronts, such as for example, path planning, self-adaptation, terrain interaction and safe locomotion. State-of-the-art methods are commonly based on a point-wise analysis of the point cloud. This analysis only considers the geometrical (e.g., normals and curvatures) as well as the robot kinematic constraints (e.g., nonholonomic, skid-steering, odometry). Unlike these methods, the CNN I’m proposing classifies the traversability of the terrain patches from both images and point clouds. The network estimates the most suitable directions for the tracked robot to autonomously crawl them. Moreover, it resorts to a CNN recognising the kind of the material of the soil (e.g., metal, concrete, stone, wood) from images as an expert network. Both the CNNs are also merged into a single network by sharing their convolutional features, so as to improve prediction accuracy.

I have been involved in the H2020-ICT-2014-1, RIA, SecondHands Project. The goal of SecondHands is to design a robot that can offer help to a maintenance technician in a pro-active manner. My main responsibility in this project is to develop a framework which allows the robot to reason about complex, either logical or statistical, structures representing its knowledge of the environment. Among others, the main novelties of this framework will be (1) the definition of such complex structures where reasoning can take place, (2) a segmentation of these structures suitable for optimising robot action planning and (3) the use of different levels of reasoning mechanisms, spanning from logical entailment to reason about time. At a preliminary step of the development of this framework I implemented a framework responsible for establishing a coherent correspondence between abstract representations and perceptual data in an embodied symbolic reasoning system. Within this framework, first the problem is transformed into a vertex coloring problem and then a solution is found, accounting for the constraints. This framework has been successfully used in conjunction with the well-known Fast-Downward algorithm for planning the actions of a humanoid robot instructed to perform a search-and-grasp-object task.