Within the realm of autonomous driving, Autonomous Racing competitions represent a well-established and unique reality: the expensive high-performance vehicles involved create a demanding proving ground for state-of-the-art algorithms, requiring high levels of precision and robustness. In this context, a new event, the Abu Dhabi Autonomous Racing League (A2RL), had its inaugural exhibition in Abu Dhabi in April 2024. This competition hosted the first head-to-head race on a road-course track, introducing teams to new and more complex challenges. Compared to oval tracks, a road-course layout poses a harder test due to the presence of multiple turns and technical sections, necessitating more intricate trajectory planning. Consequently, this complexity makes predicting the opponent’s vehicle motions non-trivial. The ability to predict opponent movements with high precision allows for better strategic planning and real-time decision-making. The various turns and the frequently changing speed require the development of a precise spatio-temporal prediction. This thesis addresses the spatio-temporal prediction problem by introducing two parallel subproblems. Firstly, it extrapolates the opponent’s racing line from the perceived data during the race, exploiting a grid-based quantization strategy and Gaussian Process Regression (GPR) to generate the final path and its associated curvature profile. Simultaneously, the maximum accelerations of the opponent are estimated using an optimization problem, which receives as input the instantaneous accelerations. Finally, the spatial and temporal information are merged to generate a predicted speed profile for the opponent. The algorithm was validated in simulation, using elaborate vehicle models that mimic the real car’s dynamics. The proposed pipeline is successful in deducing the racing line (with errors of approximately 0.25 m), the associated curvature, as well as the maximum accelerations. The predicted speed profile is also comparable with the actual opponent’s velocity.
Nel campo della guida autonoma, le competizioni tra auto da corsa autonome rappresentano una realtà ben consolidata: utilizzando veicoli ad alte prestazioni costosi, queste gare costituiscono un banco di prova unico per algoritmi all’avanguardia, richiedendo alti livelli di precisione e robustezza. In questo contesto, un nuovo evento, l’Abu Dhabi Autonomous Racing League (A2RL), ha avuto la sua gara inaugurale ad Abu Dhabi nell’Aprile 2024. Questa prima competizione ha visto realizzarsi la prima gara testa a testa su un circuito stradale, introducendo le squadre a nuove sfide più complesse . Rispetto ai circuiti ovali, un tracciato stradale è una prova più difficile a causa della presenza di molteplici curve, che necessitano di una pianificazione della traiettoria più complessa. Di conseguenza, questa complessità rende non banale anche la previsione dei movimenti dei veicoli avversari. La capacità di prevedere gli avversari con alta precisione permette una migliore pianificazione strategica e facilita le decisioni in tempo reale. Le numerose curve e le frequenti variazioni di velocità richiedono lo sviluppo di una efficace previsione spazio-temporale. Questa tesi affronta il problema della previsione spazio-temporale introducendo due sottoproblemi paralleli. In primo luogo, essa estrapola la linea dell’avversario dai dati raccolti durante la gara, utilizzando una quantizzazione basata su una griglia e Gaussian Process Regression (GPR) per generare il percorso finale e il profilo di curvatura associato. Simultaneamente, le accelerazioni massime dell’avversario vengono stimate utilizzando un problema di ottimizzazione che riceve come input le accelerazioni istantanee. Infine, le informazioni spaziali e temporali vengono fuse per generare il profilo di velocità previsto per l’avversario. La validazione dell’algoritmo è avvenuta in simulazione, utilizzando modelli veicolo accurati che riproducono le dinamiche reali. La soluzione proposta in questa tesi riesce a dedurre la linea seguita dall’avversario (con errori di circa 0.25 metri), la curvatura associata, nonché le accelerazioni massime. Il profilo di velocità predetto è comparabile con la velocità effettiva tenuta dall’avversario.
Development of an Opponent Trajectory Prediction Algorithm for High-Performance Racing Applications
Riva, Alessandro
2023/2024
Abstract
Within the realm of autonomous driving, Autonomous Racing competitions represent a well-established and unique reality: the expensive high-performance vehicles involved create a demanding proving ground for state-of-the-art algorithms, requiring high levels of precision and robustness. In this context, a new event, the Abu Dhabi Autonomous Racing League (A2RL), had its inaugural exhibition in Abu Dhabi in April 2024. This competition hosted the first head-to-head race on a road-course track, introducing teams to new and more complex challenges. Compared to oval tracks, a road-course layout poses a harder test due to the presence of multiple turns and technical sections, necessitating more intricate trajectory planning. Consequently, this complexity makes predicting the opponent’s vehicle motions non-trivial. The ability to predict opponent movements with high precision allows for better strategic planning and real-time decision-making. The various turns and the frequently changing speed require the development of a precise spatio-temporal prediction. This thesis addresses the spatio-temporal prediction problem by introducing two parallel subproblems. Firstly, it extrapolates the opponent’s racing line from the perceived data during the race, exploiting a grid-based quantization strategy and Gaussian Process Regression (GPR) to generate the final path and its associated curvature profile. Simultaneously, the maximum accelerations of the opponent are estimated using an optimization problem, which receives as input the instantaneous accelerations. Finally, the spatial and temporal information are merged to generate a predicted speed profile for the opponent. The algorithm was validated in simulation, using elaborate vehicle models that mimic the real car’s dynamics. The proposed pipeline is successful in deducing the racing line (with errors of approximately 0.25 m), the associated curvature, as well as the maximum accelerations. The predicted speed profile is also comparable with the actual opponent’s velocity.File | Dimensione | Formato | |
---|---|---|---|
2024_07_Riva_Tesi_01.pdf
non accessibile
Dimensione
9.51 MB
Formato
Adobe PDF
|
9.51 MB | Adobe PDF | Visualizza/Apri |
2024_07_Riva_Executive_Summary_01.pdf
non accessibile
Dimensione
2.33 MB
Formato
Adobe PDF
|
2.33 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/10589/223472