Master thesis - AI uncertainty to guardrails

30 credits - From AI Uncertainty to Safety Guardrails for Autonomous Heavy-Duty Vehicles 

Introduction 
 

AI-based autonomous driving increasingly relies on learned models to interpret traffic situations and propose vehicle motion. These models can perform well in familiar conditions but may fail under distribution shift. Knowing when the AI is uncertain, and how that uncertainty should affect downstream decisions, is therefore important for safe deployment. 

Background 

Learned driving components may represent uncertainty through confidence scores, probability distributions, or prediction regions. These outputs are not always calibrated, comparable, or directly useful for safety-critical decisions. 

Here, guardrails are an independent safety layer that translates rules, operating conditions, and AI uncertainty into constraints, margins, or fallback requests for verified planning and control. 

Objective 

The thesis will review how uncertainty is defined, represented, measured, and evaluated in AI-based autonomous driving, and how it can support guardrails. The focus is scene-level representations, predictions, and nominal trajectories. Sensor, perception, and localization uncertainty are included only when they propagate to these outputs. 

Job description 

The work is primarily a structured literature review, complemented where useful by lightweight computational experiments. Training a new driving policy or implementing a complete safety stack is outside the scope. The student will:

• Review uncertainty definitions and sources in learned driving-policy outputs. 

• Compare methods for estimation, representation, calibration, and evaluation. 

• Study how uncertainty can guide safety margins, constraints, interventions, and fallback decisions. 

• Recommend directions for future research on uncertainty-aware safety supervision. 

The work will be conducted with researchers and engineers at TRATON and KTH. 

Education/program/focus 

Master's student in Systems, Control and Robotics, Computer Science, Electrical Engineering, Engineering Physics, Applied Mathematics, or a related field. Experience in machine learning, probabilistic methods, autonomous driving, robotics, planning and control, or literature reviews is beneficial. 

Number of students: 1 

Start date for the thesis work: [To be agreed]