Adriano Fagiolini

In systems where many heterogeneous agents operate autonomously, with competing goals and without a centralized planner or global information repository, safety and performance can be guaranteed only by "social" rules imposed on the individual agents' behaviors. Social rules' nature is typically local, i.e. based on information made available to an agent from a small number of its neighbors. Car mobility with traffic rules and logistic robots in warehouses are canonical examples of such regulated autonomy, along with ant colonies, but many other fairly-competing autonomous systems, including distributed power plants, are coming shortly. In such systems, detecting whether any agent is not abiding to the rules is instrumental to raise an alert and take suitable countermeasures. The limited visibility due to the local nature of information however makes the misbehavior detection problem hard for any single agent, and only an exchange of information between agents may provide sufficient clues to arrive to a decision.

We consider threats that may be posed to such a society by the misbehaviors of some of its members, either due to faults or malice, and the possibility to detect and isolate them through cooperation of peers. We discuss intrusion detection algorithms, which allow detection of deviance from such rules, and algorithms to build a consensus view on the environment and on the integrity of peers, so as to improve the overall security of the society of robots. After providing a formal framework for describing social rules unifying many different applications, we study how to develop tools to automatically generate local monitors' code.

Soft robots are considered as cutting-edge technology, primarily with the vision of enabling humans a harmless physical interaction with robots. They are intrinsically endowed with dynamic modulation of elasticity while moving, which opens many opportunities to everydays' life through the achievement of human-like abilities, such as dexterity and robustness. Yet, to fully exploit their potentialities, one has to know precisely their actual stiffness, which is not a measurable quantity. In this respect, we first investigate on techniques for the estimation of stiffness and flexibility torque in robot joints. The challenge is tackled here from the motor side, but, as a pivotal point of the strategy, by considering the flexibility torque signal as an unknown input of the linear motor model. With this regard, Unknown Input Observers (UIO) are useful tools that have mainly been used for detecting system failures, by achieving correct state estimation independently of the unknown inputs.

Moreover, among soft robots, the articulated ones are such that the elasticity is concentrated at the joints which are actuated by so-called Variable Stiffness Actuators (VSA) devices. These devices are nowadays mostly provided by electrically-driven or pneumatically-driven mechanisms, that enable accurate position and velocity control, while also allowing online compliance adjustment. In this context, we study innovative solutions for adaptive control and learning both for articulated soft robots and for continuum soft robots, and compare them to traditional ones.

Continuum soft robots are endowed with morphological flexibility and compliance, and promise to have a disruptive impact, but they require facing the challenge of dealing with systems with states of theoretically infinite sizes. One possible strategy to deal with such systems is by using model-free machine learning techniques that regard a soft robot as a black box. On the other end of the spectrum, model-based techniques fully taking into account the infinite nature of the problem are still unfeasible. Within this category, we address the modeling problem and propose new, singularity- and discontinuity-free, model parameterizations overcoming issues, which establishes a link between soft-bodied and rigid-bodied robots. This connection is instrumental to the derivation of controllers which are numerically stable and well defined in the whole configuration space. We finally propose robust closed-loop position controllers for soft-bodied robots, which are based on the nonlinear adaptive control theory.

In the near future a great number of automotive applications will become possible, by leveraging on denser and faster communication networks also enabled by the 5G technology. Thanks to the Vehicle-to-Everything (V2X) architecture, vehicles, passengers, and pedestrians will be able to cooperatively plan and optimize their travel experience. They will be able to share evidence of possible hazards, including unexpected traffic jams in tunnels, road damages, anomalous behavior of human drivers and autonomous pilots, thus improving the overall safety of passengers and pedestrians.

In this scenario, the race towards (electric) vehicles with full self-driving capacity has just begun. However, several obstacles have to be overcome before this technology goes mainstream to the market, including infrastructure modernization, legislations definition, and stronger guarantees on the ability of an autonomous vehicle to detect and react to uncertainties caused by unexpected changes in the driving conditions. A notable example is the field of self-driving vehicles, where the DARPA Grand Challenge and Urban Challenge have pushed the robotics community to build autonomous cars for unstructured or urban scenarios. We investigate on and propose fast estimators) of the environmental conditions (road, wind, etc.) and suitable longitudinal and lateral controllers that can promptly react and ensure safety. We believe this research can strengthen the applicability of self-driving solutions and promote their usage in the society for safer roads.

Unmanned Aerial Vehicles (UAV) have been drawing increasing attention for more than two decades in various application fields, ranging from the industry to the military and from the service to the entertainment. Due to their ability to reach places, particularly those that are hardly accessible by land vehicles, UAVs are convenient tools for monitoring areas where natural disasters have just occurred. This is of high interest in the Pacific region, where remote neighborhoods need to be rapidly checked after cyclones or floods. The U.N. Food and Agriculture Organization (FAO) has launched in the Philippines a drone initiative to assess where agricultural land is at most risk of natural disasters and how to rapidly evaluate damages after they occur. It is strongly believed that the adoption of UAV platforms can significantly enhance risk and damage assessments, but also revolutionize the way to prepare for and respond to disasters. Using this kind of aircraft in hostile conditions, including strong wind gusts, is still an open problem. The nonlinearity of the system model, along with its underactuation, must taken into account, as they otherwise negatively affect on the mission performance, which is particularly true for lightweight low-cost quadrotors.

In this context, we study solutions that are applicable to low-cost multirotor aircraft, which allows avoiding direct wind speed measurement via anemometers. We seek solutions with the appealing features of being simple, having low computation cost, being able to obtain a fast response to wind gusts, and implementable on virtually all aircraft systems, as a stand-alone solution or an extension plugin for existing controllers. Along this line, we propose an innovative approach where wind disturbance is modeled as an unknown exogenous input and then it is estimated via an Unknown Input-State Observer (UIO). In order to further improve the promptness and efficacy of the controlled aircraft, we describe alternative solutions using Nonlinear UIOs and model-decoupling, ESO-based techniques which are also robust to model uncertainty.

A problem of great interest in real applications using large power machines and motors is reducing the number of sensors needed for processing a given control law. Motion control of systems with induction motors (IM) without speed sensor (sensorless) has been longly addressed by many authors, since these systems often operate in unaccessible environments. A crucial problem to solve for the implementation of sensorless control laws is the determination of both the rotor flux vector and the speed.

We investigate on new solutions to estimate the state of these machines and to control them despite the uncertainties of the model, with a special focus on low-speed operating conditions and with varying load conditions. As is well known, the observability property of the model is crucial for the existence of state observers, a property that is known to be lost at zero rotor speed. We provide a new approach to estimate the speed of an IM by using an extended Kalman filter (EKF), which remains valid at very low rotor speed. Also, we reformulate the motor model by using complex-valued variables, which allows reducing the size of the model as well as simplifying the observability property. This enables the derivation of Extended Complex-valued Kalman Filters (ECKF) whose main feature is a smaller required computational time.