Corner Case Detection for Autonomous Driving using World Model

Institut

Lehrstuhl für Fahrzeugtechnik (TUM-ED)

Typ

Semesterarbeit / Masterarbeit /

Inhalt

experimentell / theoretisch / konstruktiv /  

Beschreibung

Introduction

With the advancement of autonomous driving, many autonomous driving (AD) stacks are aiming at SAE level 4. However, even level 4 AD systems still encounter corner cases (CCs) and will disengage once they cannot solve the situation. To tackle these CCs will be an essential problem in autonomous driving. Therefore, it is crucial for the continuous development of AD systems to detect, collect, and eventually solve these CCs. In the light of this, world models provide a possibility to detect and to predict these CCs by predicting the near future based on current and previous vehicle states and actions.

 

Description

In the project, you will develop a framework for corner case detection in AD using existing world models. Furthermore, you will investigate the performance of the world models based on their input/output modalities and prediction horizons.

The project can be described with the following tasks:

• Develop a framework to integrate existing world models for corner case detection in AD
• Finetune the existing world models with simulation and real-world data
• Evaluate the performance of the world models for CC detection

Voraussetzungen

Prerequisites

• Interest and selfmotivation in the topic
• Handson experience with Python, Pytorch, Tensorflow
• (Preferred) Knowledge and/or previous experience with diffusion models
• (Preferred) Knowledge and previous experience with large model finetuning with multi-GPU setup

Tags

FTM Studienarbeit, FTM AV, FTM AV Safe Operation, FTM Su, FTM Informatik

Möglicher Beginn

sofort

Kontakt

Xiyan Su, M.Sc.
Raum: MW3507
Tel.: +49 89 289 15340
xiyan.sutum.de

Ausschreibung