The study and development of Autonomous Driving Systems (ADS) have seen strong growth in recent years. Alongside the academic field, many private car are allocating huge investments to favor the advancement of these technologies, increasing road safety, and limiting human errors. Furthermore, in recent years, the first self-driving vehicle sports competitions have been organized, with the aim of accelerating progress in the sector and achieving significant results in the field of multi-agent autonomous driving. In particular, the main competitions are two: the Indy Autonomous Challenge and Roborace, and the MOVE research team, which assisted in the realization of this project, participates in both. The main purpose of this thesis is in fact to design and develop planning and estimation algorithms for AS.CAR.I. (AutonomouS CAR Intelligence), the autonomous driving pilot of the PoliMOVE Autonomous Racing Team participating in the Season Beta of the Roborace competition. In details, this thesis analyzes possible solutions to infer information on the opponent vehicles present on the track and to interact with them: in the first instance, an algorithm is shown for the prediction of the trajectory of these vehicles based on the knowledge of their Cartesian coordinates, of their orientation and of the race lines that these agents can potentially run. Next, a method for deriving the longitudinal and lateral acceleration profiles of the adversary vehicles is illustrated, in this case starting from a knowledge base that includes the vehicle speed module and its orientation. Then, the obtained profiles will be used for the development of an algorithm for the estimation of the limit ellipse of the opponent's performance, based on the solving of a Least Square optimization problem and on an additional contribution in open loop. Our planner needs these limits to predict accurately the speed profile of the opponent, along its estimated future trajectory. Then, such information will be exploited to make decisions on the feasibility of overtaking maneuvers toward the said opponent. Finally, the last part of the thesis deals with discussing two possible implementations of vehicle-to-vehicle distance controllers. This distance, as it will be shown in the thesis, is determined starting from the knowledge of the Cartesian coordinates, the orientation, and the dimensions of the vehicles on the track. The first proposed controller is characterized by a component proportional to the tracking error of the reference distance and one proportional to the difference in the speeds of the two vehicles, while the second also introduces a double derivative contribution - proportional to the difference in accelerations - to overcome the limitations of the former in driving situations with high acceleration variations, as often happens in a racing context. The whole work is based on a first phase of algorithm design, followed by the implementation of the proposed solutions and then by an intense phase of experimental analysis both in the simulation environment and on the physical vehicle. The obtained results demonstrate the effectiveness of the proposed solutions.
Lo studio e lo sviluppo di Autonomous Driving Systems (ADS) ha visto una forte crescita negli ultimi anni. Accanto all'ambito accademico, molte compagnie automobilistiche private hanno deciso di stanziare ingenti investimenti per favorire l'avanzamento di queste tecnologie, con lo scopo di aumentare la sicurezza stradale e limitare gli errori umani. Negli ultimi anni inoltre, sono nate le prime competizioni sportive di veicoli a guida autonoma, con lo scopo di accelerare il progresso in materia e raggiungere risultati significativi nell'ambito della guida autonoma multi-agente. In particolare, la principali competizioni sono due: la Indy Autonomous Challenge e Roborace, ed il team di ricerca MOVE, che ha coadiuvato la realizzazione di questo elaborato, partecipa ad entrambe. Lo scopo principale di questa tesi è infatti quello di progettare e sviluppare algoritmi di pianificazione e di stima per AS.CAR.I. (AutonomouS CAR Intelligence), l'AI Driver del PoliMOVE Autonomous Racing Team che partecipa alla Season Beta della competizione Roborace. In detaglio, questa tesi analizza possibili soluzioni per inferire informazioni sui veicoli avversari presenti sul tracciato e sull'interazione con questi: in prima istanza, viene mostrato un algoritmo per la predizione della traiettoria di un avversario basato sulla conoscenza delle sue coordinate cartesiane, del suo orientamento e delle linee gara che tale veicolo può potenzialmente percorrere. In seguito, viene illustrato un metodo per derivare i profili di accelerazione longitudinale e laterale degli avversari, in questo caso partendo da una base di conoscenza che comprende il modulo della velocità del veicolo ed il suo orientamento. Tali profili ottenuti saranno quindi essenziali per lo sviluppo di un algoritmo per la stima dell'ellisse limite delle performance dell'avversario, basato sulla risoluzione di un problema di ottimizzazione Least Square e su un contributo addizionale in anello aperto. Tali limiti sono infatti necessari per realizzare un profilo di velocità accurato dell'avversario lungo la sua traiettoria futura stimata, informazione che verrà poi sfruttata dal Driver per prendere decisioni sulla fattibilità di manovre di sorpasso nei confronti degli altri veicoli. Infine, l'ultima parte della tesi si occupa di discutere due possibili implementazioni di controllori di distanza veicolo-veicolo. Tale distanza, come verrà mostrato dall'elaborato, viene determinata a partire dalla conoscenza delle coordinate cartesiane dei veicoli sul tracciato, del loro orientamento e delle loro dimensioni. Il primo dei controllori proposti è caratterizzato da una componente proporzionale all'errore di tracking della distanza di riferimento ed uno proporzionale alla differenza delle velocità dei due veicoli, mentre il secondo introduce anche un contributo doppiamente derivativo - proporzionale cioè alla differenza delle accelerazioni - per superare le limitazioni del primo in situazioni di guida con elevate variazioni di accelerazione, come spesso accade in un contesto di gara. L'intero lavoro si basa su una prima fase di design degli algoritmi, seguita dall'attività di implementazione delle soluzioni proposte e quindi da un'intensa fase di analisi sperimentale sia in ambiente di simulazione che sul veicolo fisico. I risultati ottenuti hanno dimostrato l'efficacia delle soluzioni proposte.
Design and development of estimation and planning algorithms for autonomous racing
Anselmi, Edoardo
2020/2021
Abstract
The study and development of Autonomous Driving Systems (ADS) have seen strong growth in recent years. Alongside the academic field, many private car are allocating huge investments to favor the advancement of these technologies, increasing road safety, and limiting human errors. Furthermore, in recent years, the first self-driving vehicle sports competitions have been organized, with the aim of accelerating progress in the sector and achieving significant results in the field of multi-agent autonomous driving. In particular, the main competitions are two: the Indy Autonomous Challenge and Roborace, and the MOVE research team, which assisted in the realization of this project, participates in both. The main purpose of this thesis is in fact to design and develop planning and estimation algorithms for AS.CAR.I. (AutonomouS CAR Intelligence), the autonomous driving pilot of the PoliMOVE Autonomous Racing Team participating in the Season Beta of the Roborace competition. In details, this thesis analyzes possible solutions to infer information on the opponent vehicles present on the track and to interact with them: in the first instance, an algorithm is shown for the prediction of the trajectory of these vehicles based on the knowledge of their Cartesian coordinates, of their orientation and of the race lines that these agents can potentially run. Next, a method for deriving the longitudinal and lateral acceleration profiles of the adversary vehicles is illustrated, in this case starting from a knowledge base that includes the vehicle speed module and its orientation. Then, the obtained profiles will be used for the development of an algorithm for the estimation of the limit ellipse of the opponent's performance, based on the solving of a Least Square optimization problem and on an additional contribution in open loop. Our planner needs these limits to predict accurately the speed profile of the opponent, along its estimated future trajectory. Then, such information will be exploited to make decisions on the feasibility of overtaking maneuvers toward the said opponent. Finally, the last part of the thesis deals with discussing two possible implementations of vehicle-to-vehicle distance controllers. This distance, as it will be shown in the thesis, is determined starting from the knowledge of the Cartesian coordinates, the orientation, and the dimensions of the vehicles on the track. The first proposed controller is characterized by a component proportional to the tracking error of the reference distance and one proportional to the difference in the speeds of the two vehicles, while the second also introduces a double derivative contribution - proportional to the difference in accelerations - to overcome the limitations of the former in driving situations with high acceleration variations, as often happens in a racing context. The whole work is based on a first phase of algorithm design, followed by the implementation of the proposed solutions and then by an intense phase of experimental analysis both in the simulation environment and on the physical vehicle. The obtained results demonstrate the effectiveness of the proposed solutions.| File | Dimensione | Formato | |
|---|---|---|---|
|
Thesis_EdoardoAnselmi.pdf
non accessibile
Dimensione
13.14 MB
Formato
Adobe PDF
|
13.14 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/183634