Online Adaptation of autonomous robots
Online version • Manuscript • Slides
Abstract
Robots diffusion is constantly increasing, yet their applications remain largely confined to industries and other highly controlled environments; the cause is their inability to address unexpected situations that can occur in the open world. We argue that endowing robots with the ability to adjust their own behavior can improve their operational efficacy in uncontrolled environments. In this dissertation, we investigate factors enabling the runtime modification of robot behaviors. The proposed operative approach consists in the systematic investigation and resolution of increasingly complex conditions. During the exploration, we employ robots with minimal computational capabilities, so as to investigate the factors enabling adaptation without them being obscured by the complexity of the control system employed; this also permits providing adaptive capabilities to simple robots, which might not be able to operate complex control software.
The principal contributions of this dissertation include the identification of factors essential for adapting systems, among which stand out the need for self-evaluation, re-evaluation, and time- and intensity-modulated adaptation. Additionally, we highlight the importance of well-devised evaluation criteria and of their inherent characteristics, mainly the evaluation time and whether they tend to induce the emergence of a specific behavior or permit a broad spectrum of outcomes. We propose adaptive mechanisms capable of operating on various robot control architectures, including one designed offline by an automatic system with the explicit intent of optimizing its runtime adaptive capabilities. Finally, we present, define, and propose solutions to the problem of “situationality”, a condition in which external factors determine the performance of the robot regardless of its capabilities.
Overall, this dissertation aims to contribute to the literature on online adaptation in robotics, shedding light on some frequently underestimated issues. We expect that the proposed solutions and considerations will support the future advancement of this discipline.
@phdthesis{baldini2026online,
title = {Online Adaptation of autonomous robots},
author = {Baldini, Paolo},
year = {2026},
school = {University of Bologna},
note = {Supervisor: Roli, Andrea, Co-Supervisor: Viroli, Mirko}
}