Malte Stelzer, M.Sc.

Offene Abschlussarbeiten

Time-Variant L1-Loss Weighting for Event-Aware End-to-End Autonomous Driving
This work extends L1-loss in imitation learning by introducing time-dependent weighting for critical traffic events (e.g., turns, lane changes, braking). Lateral and longitudinal deviations are treated separately to better reflect safety and driving dynamics. Adaptive weighting based on context factors (traffic density, speed, road layout) emphasizes relevant scenarios, while regularization and curriculum strategies address potential instability. The aim is more fine-grained error evaluation and improved model behavior in safety-critical situations.

Temporal Context Integration for Anticipatory End-to-End Driving Decisions
Current E2E architectures often process sensory input frame by frame, limiting their ability to anticipate dynamic situations. This project investigates temporal fusion techniques (e.g., RNNs, Transformers, sequence encodings) to explicitly model sequential dependencies. The goal is to enhance trajectory consistency, improve proactive decision-making, and assess the contribution of temporal context to safer, more stable driving policies.

Explainability via BEV-Based Auxiliary Outputs in End-to-End Trajectory Planning
To increase interpretability of E2E models, this work integrates bird’s-eye-view (BEV) auxiliary outputs—such as map features or object positions—into trajectory planning. Building on architectures like TransFuser, structured intermediate representations are explored not only as training aids but as explainable decision bases. The objective is to make planning logic transparent in complex scenarios, supporting developer insight and the integration of safety-critical control mechanisms.

Interactive Visualization and Analysis Tool for CARLA Datasets and Autonomous Driving Models
This project extends existing visualization tools for CARLA by integrating simulation control, dataset exploration, and model behavior analysis into a unified interface. Users can dynamically replay scenarios, inspect trajectories, and manipulate simulation parameters to probe model decisions. Additionally, dataset-level analytics (e.g., distribution of traffic events, corner cases, failure modes) are incorporated to identify strengths and weaknesses in training data and learned policies.