Ensuring safety in autonomous driving requires not only accurate predictions of surrounding traffic agents but also reliable quantification of the uncertainty associated with those predictions. This thesis addresses the problem of rigorous uncertainty quantification in trajectory forecasting for autonomous heavy vehicle combinations operating in highway environments. In such safety-critical scenarios, black-box machine learning models often suffer from performance degradation due to out-of-distribution inputs, which can lead to unsafe path planning decisions if not properly accounted for. To tackle this, we investigate and extend a family of methods known as conformal prediction: a model-agnostic framework that wraps around any trajectory predictor and constructs prediction regions with guaranteed statistical coverage, without assuming specific data distributions. The project analyzes state-of-the-art conformal methods for both finite and infinitehorizon control settings, with focus on their empirical performance in simulated highway scenarios. We propose novel extensions that embed heuristic uncertainty or leverage timeseries modeling to adapt prediction sets online, improving coverage under distribution shifts. These methods are tested in a closed-loop autonomous truck simulator, where predicted regions are integrated with a model predictive controller (MPC) responsible for generating safe and feasible paths. This MPC integration is not the focus of the research, but is instead used as a case study to demonstrate how quantified uncertainty can inform downstream planning. Results show that the novel adaptive conformal methods maintain coverage under challenging and changing traffic configurations and produce uncertainty bounds that are more reliable than those based on heuristic methods. This work enables downstream systems, such as path planners, to better handle prediction uncertainty with formal guarantees, closing the gap between learning-based forecasting and robust motion planning.
Garantire la sicurezza nella guida autonoma richiede non solo una previsione accurata del comportamento dei veicoli circostanti, ma anche una stima affidabile dell’incertezza associata a tali previsioni. Questa tesi affronta il problema della quantificazione rigorosa dell’incertezza nella previsione delle traiettorie per veicoli pesanti autonomi operanti in scenari autostradali. In contesti così critici, i modelli di Machine Learning tendono a perdere affidabilità quando vengono utilizzati in scenari diversi da quelli osservati durante la fase di addestramento, il che può compromettere la sicurezza della pianificazione della traiettoria se l’incertezza non viene adeguatamente considerata. Per affrontare questo problema, vengono analizzati e potenziati metodi appartenenti alla famiglia della Conformal Prediction (CP): un approccio agnostico rispetto al modello predittivo, capace di fornire regioni predittive con garanzie statistiche di copertura, indipendentemente dalla distribuzione dei dati. Il lavoro si concentra sull’analisi empirica di metodi avanzati di CP, sia per orizzonti temporali finiti che infiniti, valutandone le prestazioni in scenari autostradali simulati. Sono inoltre proposte estensioni di tali metodi che incorporano euristiche o modelli autoregressivi per adattare dinamicamente le regioni predittive, migliorando la copertura in condizioni di Distribution Shifts. Tali metodi vengono integrati in un simulatore closed-loop per la guida autonoma di camion, dove le regioni predittive vengono utilizzate all’interno di un Model Predictive Control (MPC) per generare traiettorie sicure. L’integrazione con l’MPC non rappresenta il focus principale del lavoro, ma serve da caso studio per mostrare come una quantificazione dell’incertezza possa benificiare la pianificazione della traiettoria. I risultati dimostrano che i metodi CP adattivi proposti mantengono buoni livelli di copertura anche in condizioni di traffico complesse e dinamiche, producendo regioni di predizione più affidabili rispetto a quelli ottenuti con approcci euristici. Il lavoro fornisce un ponte tra modelli predittivi data-driven e pianificazione di traiettoria robusta, ponendo le basi per un’integrazione più sicura all’interno dei sistemi di guida autonoma.
Robust trajectory forecasting for autonomous vehicles using conformal machine learning
LODETTI, MARCO
2024/2025
Abstract
Ensuring safety in autonomous driving requires not only accurate predictions of surrounding traffic agents but also reliable quantification of the uncertainty associated with those predictions. This thesis addresses the problem of rigorous uncertainty quantification in trajectory forecasting for autonomous heavy vehicle combinations operating in highway environments. In such safety-critical scenarios, black-box machine learning models often suffer from performance degradation due to out-of-distribution inputs, which can lead to unsafe path planning decisions if not properly accounted for. To tackle this, we investigate and extend a family of methods known as conformal prediction: a model-agnostic framework that wraps around any trajectory predictor and constructs prediction regions with guaranteed statistical coverage, without assuming specific data distributions. The project analyzes state-of-the-art conformal methods for both finite and infinitehorizon control settings, with focus on their empirical performance in simulated highway scenarios. We propose novel extensions that embed heuristic uncertainty or leverage timeseries modeling to adapt prediction sets online, improving coverage under distribution shifts. These methods are tested in a closed-loop autonomous truck simulator, where predicted regions are integrated with a model predictive controller (MPC) responsible for generating safe and feasible paths. This MPC integration is not the focus of the research, but is instead used as a case study to demonstrate how quantified uncertainty can inform downstream planning. Results show that the novel adaptive conformal methods maintain coverage under challenging and changing traffic configurations and produce uncertainty bounds that are more reliable than those based on heuristic methods. This work enables downstream systems, such as path planners, to better handle prediction uncertainty with formal guarantees, closing the gap between learning-based forecasting and robust motion planning.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_06_24_Lodetti_Tesi.pdf
accessibile in internet per tutti
Descrizione: Tesi Lodetti Marco
Dimensione
6.87 MB
Formato
Adobe PDF
|
6.87 MB | Adobe PDF | Visualizza/Apri |
|
2025_06_26_Lodetti_executive_summary.pdf
accessibile in internet per tutti
Descrizione: Executive Summary Lodetti Marco
Dimensione
1.13 MB
Formato
Adobe PDF
|
1.13 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/10589/239617