Every open topic says what you would build and names the person to write to.
Take a vectorised HD map of a real Munich intersection and turn it into a CARLA town that HABIT can run scenarios in, so we can evaluate on the places we actually record.
Python, CARLA, some 3D tooling. Good if you like making things fit together.
SatMap uses satellite imagery as a prior, LIE enhances LiDAR intensity. Combine the two: use the scanner where the imagery is stale and the imagery where the scanner is blind, and measure what each contributes.
PyTorch, 3D geometry, willingness to work with nuScenes and our own recordings.
Diffusion policies and language-model planners both produce trajectories, and they fail in different ways. Implement both in our closed-loop CARLA setting and find out where each breaks.
PyTorch, CARLA or a willingness to learn it, curiosity about generative models.
End-to-end planners produce a trajectory and no reason for it. Define a measurable score for how well a model’s stated reason matches what it actually did, and evaluate it on BEVDriver.
Python, some familiarity with deep learning. An interest in evaluation design matters more than experience.
A recording day produces hours of LiDAR and camera data and almost no labels. Build the pipeline that turns a raw drive into training data: pre-annotate with existing detectors, track through occlusions, and surface only the frames a human still has to check.
Python, PyTorch, patience with real data. ROS 2 helps but can be learned here.
If none of these fit but the field does, write to the lab head with what you would like to work on.
| Topic | Kind | Supervisor | Since |
|---|---|---|---|
| Scenario generation from aerial recordingsTurning DeepUrban's aerial interactions into replayable CARLA scenarios. | BA | Constantin Selzer | 2026-06 |
| Explainability probes for BEVDriverTesting whether BEVDriver's stated reasons correspond to the features it actually uses. | BA | Katharina Winter | 2026-08 |
| Occlusion-aware fusion of satellite priorsDeciding per map element whether the prior or the live sensor should be trusted. | MA | Kanak Mazumder | 2026-05 |
| Cyclist intention from 3D pose sequencesBuilding on bikeActions: predicting a cyclist's next manoeuvre from a sequence of 3D poses rather than from the trajectory alone. | MA | Max A. Büttner | 2026-04 |
| Topic | Kind | Supervisor | Finished | What came of it |
|---|---|---|---|---|
| Human action scenarios in CARLA | MA | Mohan Ramesh | 2025-10 | Fed into HABIT, WACV 2026. |
| Open-vocabulary retrieval of driving motionStefan Englmeier | MA | Max A. Büttner | 2025-03 | Published as ContextMotionCLIP; the author is now a PhD candidate in the lab. |
| Action labels for cyclist recordingsLuca Köcher, Mario Finkbeiner and Sebastian Niebler | MA | Max A. Büttner | 2025-12 | Became bikeActions, ICPR 2026 — all three are co-authors. |
| Scene captioning from multimodal BEV mapsFelix Brandstätter | MA | Katharina Winter | 2025-04 | Published as BEV-LLM with the student as first author. |
