Dottorato in INGEGNERIA INFORMATICA

Titolo della tesi: Semantic Representations for Effective Robot Behaviors in Unstructured Environments

In recent years, advancements in robotics have led to remarkable progress in the development of autonomous systems capable of navigating and interacting with their environment. One fundamental aspect of robotic autonomy lies in the ability of the robot to comprehend its surroundings and make informed decisions based on the available information. This Ph.D. thesis aims at enhancing the robots’ autonomy by building richer representations of the surroundings and making the agents capable of dynamically adjusting their behaviors accordingly. By taking advantage of a broad spectrum of sensory inputs, semantic representations are also presented and evaluated to enrich the traditional sensory inputs. In this journey to improve robots’ autonomy, we present three main steps: creating a rich environment representation to improve robot behaviors, building a semantic layer on top of it, and, finally, creating benchmarking environments to evaluate the robot’s environmental awareness. To address the first goal, a blend of machine learning algorithms and model-based approaches are used to integrate secondary perceptions into robot representation to improve autonomous robots’ behavior. This integration aims at enhancing robots’ navigational and interactive capabilities, showing benefits on mobile robots, wheeled, and humanoids, starting with single robots and extending to multi-robot systems. The second part of the thesis takes a step further introducing semantic representations. It emphasizes the importance of semantic understanding in robotics, particularly for enhanced navigation, interaction, autonomy, and human-robot interaction. The research builds upon semantic maps, which enhance traditional geometric mapping with semantic information, allowing robots to discern not just the presence of objects, but also their properties and functionalities. This semantic layer is crucial for informed navigation strategies and context-aware decision-making. The third part presents benchmarking in controlled environments, such as RoboCup and SciRoc competitions, to evaluate the effectiveness of the above mentioned approaches. An excellent example is represented by the design of an episode of the SciRoc - Smart Cities competition. These real-world testing grounds illustrate how robots, equipped with enhanced perceptual and semantic capabilities, can properly navigate complex environments and interact intelligently with other robots, humans, and their surroundings. Novel metrics are presented and exploited for evaluating the semantic representation of the environment. In summary, this thesis enhances the perceptual capabilities and situational awareness of autonomous robots. It contributes with novel approaches showcasing a range of applications where robots can automatically adapt their behavior in response to the variations of their environments. It introduces methodologies for semantic mapping, decision-making systems adaptable to various environments, and for improving human-robot interactions in shared operational scenarios.