Rate my driving: Developing an AI Framework for the Multi-Dimensional Assessment of Human-Likeness in Driving Behavior

Institut

Professur für autonome Fahrzeugsysteme (TUM-ED)

Typ

Semesterarbeit / Masterarbeit /

Inhalt

experimentell / theoretisch /  

Beschreibung

Motivation

The development of autonomous driving is advancing rapidly. In the field of end-to-end approaches, in particular, new methods are regularly published that achieve high scores on datasets and benchmarks. Currently, the primary focus of optimization is on functional safety, traffic rule compliance, and mastery of safety-critical edge cases.

However, a critical and largely unresolved question remains: HOW does the vehicle drive? Yet, at a time when (semi-)autonomous vehicles and human drivers share the road, improving the safety of autonomous vehicles does not necessarily go hand in hand with increased acceptance among human road users. Driving behavior that appears unpredictable, abrupt, or generally “non-human” can lead to uncertainty and a lack of trust among human road users, even if it is formally safe. It may therefore be essential for the development of autonomous vehicles to teach them a kind of social intelligence or “roadmanship.” 

Traditional evaluation metrics, which compare the behavior of autonomous vehicles against predefined ground truth, fail to adequately account for these aspects. Human driving is more than can be captured by a “ground truth” solution: it is inherently multimodal. For human drivers, there are often a variety of plausible yet safe alternative courses of action in most situations. Therefore, there is a need for an evaluation framework that goes beyond classic geometric/kinematic considerations and enables a semantic and cognitive assessment of driving behavior.

The goal of this thesis is to develop an AI-based framework that acts as an "intelligent observer," continuously monitoring and assessing the human-likeness of autonomous agents.

Work packages

• Literature research for the evaluation of “human-likeness” in driving behavior
• Development and curation of a dataset
• Implementation and training of the AI framework
• Benchmarking the chosen approach
• Written documentation of the procedure and results

Voraussetzungen

Requirements

Must-have

• Independent and organized working style
• Programming experience (preferably Python)
• Curious mindset
• Open-minded team player

Nice-to-have

• Knowledge in Autonomous Driving and/or Robotics
• Familiarity with frameworks such as PyTorch or TensorFlow
• Experience with training and deploying AI models

 

Ready to bring a human touch to autonomous driving? The work can start immediately. The work can be conducted in English or in German.
Please send your current transcript of records, your CV, and a brief description of your motivation to christian.oefinger@tum.de.

Tags

AVS Oefinger

Möglicher Beginn

sofort

Kontakt

Christian Oefinger
christian.oefingertum.de