Autonomous driving is rapidly evolving from controlled scenarios to complex urban environments. In recent years, autonomous racing has emerged as a crucial catalyst in this progression. By pushing vehicles to operate at the physical limits of handling, autonomous racing provides an unparalleled testbed for developing and validating high-performance perception, planning, and control strategies. Insights gained from these extreme conditions directly inform the design of safer, more efficient, and more reliable autonomous systems for public roads. As a result, autonomous racing serves not merely as a competitive domain but as a strategic accelerator for future mobility, advancing the technologies that will underpin the next generation of autonomous vehicles. A prominent event in this domain is the Abu Dhabi Autonomous Racing League (A2RL). This competition challenges teams from all over the world to compete in a series of autonomous challenge, including multi-vehicle races on the iconic Yas Marina circuit. In this context, this thesis focuses on the development of a local planning algorithm for the online generation of time-optimal merging segments that allow the autonomous vehicle to smoothly transition between two trajectories provided by a higher-level planner. This work formulates a novel quadratic Optimal Control Problem (OCP) in Frenet reference frame. The cost function of the proposed OCP minimizes not only the curvature of the merging path but also its spatial derivative. This dissertation underscores the critical role of the latter term in ensuring that the resulting reference trajectory remains trackable by the underlying controllers at the limits of vehicle handling. Extensive simulations conducted on the Yas Marina circuit demonstrate drastic improvements in terms of smoothness of planned merging maneuvers. Compared to a benchmark algorithm, the proposed methodology increases the probability of generating trackable merging maneuvers from 24.76% to 93.5%, all while preserving time competitiveness.
Il mondo della guida autonoma evolve sempre più rapidamente, affrontando scenari sempre più complessi come gli intricati contesti urbani. In questo senso, le competizioni a guida autonoma stanno fungendo da importanti catalizzatori: spingendo le macchine fino al limite, queste competizioni rappresentano un banco di prova eccezionale per sviluppare e validare algoritmi di percezione, planning e controllo ad alte prestazioni. La competenza guadagnata in queste condizioni estreme traina la crescita di sistemi autonomi più sicuri e affidabili per la strada pubblica. Per questo motivo, le competizioni a guida autonoma non rappresentano semplicemente uno sport spettacolare, ma fungono da acceleratori per la mobilità del futuro, facendo avanzare le tecnologie su cui si poggierà la prossima generazione di veicoli a guida autonoma. Un evento di rilievo in questo settore è l'Abu Dhabi Autonomous Racing League (A2RL). In questa competizione, team da tutto il mondo si confrontano in una serie di sfide a guida autonoma che comprende anche gare multiveicolo sull'iconico tracciato di Yas Marina. In questo contesto, questa tesi presenta un algoritmo di planning per la generazione online di percorsi che permettano ad un veicolo a guida autonoma di passare da una traiettoria di partenza a una di arrivo, indicata da un livello di planning sovrastante. Questo lavoro formula un nuovo problema di ottimizzazione quadratico nel sistema di riferimento di Frenet. In particolare, la funzione di costo del problema di ottimizzazione non penalizza solo la curvatura del percorso, ma anche la derivata spaziale della curvatura. Questa dissertazione evidenzia l'importanza di quest'ultima nel generare manovre che siano seguibili dal sottostante controllore. Simulazioni condotte sul tracciato di Yas Marina dimostrano il drastico miglioramento raggiunto in termini di fluidità delle manovre. Confrontato con un algoritmo di riferimento, la metodologia presentata aumenta la percentuale di manovre generate che possono essere seguite correttamente dal controllore, dal 24.76% al 93.5%, mantenendo la competitività delle stesse.
Development of a local planning algorithm for smooth merging segments in autonomous racing
Farinella, Giacomo
2024/2025
Abstract
Autonomous driving is rapidly evolving from controlled scenarios to complex urban environments. In recent years, autonomous racing has emerged as a crucial catalyst in this progression. By pushing vehicles to operate at the physical limits of handling, autonomous racing provides an unparalleled testbed for developing and validating high-performance perception, planning, and control strategies. Insights gained from these extreme conditions directly inform the design of safer, more efficient, and more reliable autonomous systems for public roads. As a result, autonomous racing serves not merely as a competitive domain but as a strategic accelerator for future mobility, advancing the technologies that will underpin the next generation of autonomous vehicles. A prominent event in this domain is the Abu Dhabi Autonomous Racing League (A2RL). This competition challenges teams from all over the world to compete in a series of autonomous challenge, including multi-vehicle races on the iconic Yas Marina circuit. In this context, this thesis focuses on the development of a local planning algorithm for the online generation of time-optimal merging segments that allow the autonomous vehicle to smoothly transition between two trajectories provided by a higher-level planner. This work formulates a novel quadratic Optimal Control Problem (OCP) in Frenet reference frame. The cost function of the proposed OCP minimizes not only the curvature of the merging path but also its spatial derivative. This dissertation underscores the critical role of the latter term in ensuring that the resulting reference trajectory remains trackable by the underlying controllers at the limits of vehicle handling. Extensive simulations conducted on the Yas Marina circuit demonstrate drastic improvements in terms of smoothness of planned merging maneuvers. Compared to a benchmark algorithm, the proposed methodology increases the probability of generating trackable merging maneuvers from 24.76% to 93.5%, all while preserving time competitiveness.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Farinella_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary
Dimensione
836.76 kB
Formato
Adobe PDF
|
836.76 kB | Adobe PDF | Visualizza/Apri |
|
2025_12_Farinella_Tesi.pdf
non accessibile
Descrizione: Tesi
Dimensione
6.05 MB
Formato
Adobe PDF
|
6.05 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/10589/247054