Master thesis - Efficient World Representations for End-to-End Autonomous Driving

30 hp - Efficient World Representations for End-to-End Autonomous Driving 

Introduction
 

A Master's thesis is an excellent way to get closer to TRATON Group R&D and build relationships for the future. In this thesis, you will contribute to EGoPT, a new industrial research project on World Models for Autonomous Driving. 

Background 

Modern autonomous vehicles generate large amounts of sensor data from cameras, lidar, radar, and vehicle-state signals. Processing all this information at full resolution and over long temporal histories can exceed the computational budget available for real-time planning. Compressing it too aggressively, however, may remove information that is important for planning or safety. 

The VINNOVA-FFI EGoPT project investigates how compact, task-aligned representations of multimodal sensor data can support real-time trajectory planning for autonomous heavy-duty vehicles. Current end-to-end driving methods and evaluation tools are mainly developed for passenger cars, while trucks introduce additional constraints related to vehicle size, articulation, load-dependent dynamics, braking distance, and computational resources. 

The thesis will primarily use public datasets, open-source models, simulation, and planning benchmarks and emerging truck-focused resources. 

Objective 

The thesis will investigate an initial research question within EGoPT. Possible directions include:

• Efficient spatial or temporal representations: compress sensor observations or scene history while preserving information needed for planning. 

• Adaptive representations: allocate a limited token or compute budget to the cameras, regions, or information most relevant to the current driving situation. 

• Safety-aware compression: evaluate whether compact representations preserve safety-critical information. 

• Heavy-duty vehicle generalization and evaluation: adapt learned planners to different vehicle configurations, or evaluate and extend public benchmarks with articulated or configuration-dependent constraints. 

Job description 

During the thesis period, you will: 

• Review relevant literature and help define a focused research question. 

• Set up or reproduce an open-source autonomous-driving model, simulator, or benchmark. 

• Implement and evaluate a method, benchmark extension, or experimental framework. 

• Analyze relevant trade-offs in planning performance, safety, generalization, latency, memory, or computational cost. 

• Document the work, present the results, and write the final thesis report.