Teaching

GNSS-free localization of an autonomous vehicle by triangulating distances to nearby cities from roadside milestone boards, combining multilateration with vehicle odometry and evaluating on data from the lab's Lexus platform.

Making the lab's offline map-based particle filter localisation pipelines (milestone boards on the highway scale, street-name signs on the city scale) run in real time on board the lab's Lexus RX450h, with on-road evaluation in shadow mode.

Investigating whether odometry error of an autonomous vehicle can be predicted from real-world driving data using machine learning, comparing models on prediction accuracy, computational cost and energy consumption — including routes and driving conditions not seen during training — with a view to real-time use.

Building a pipeline for reconstructing and rendering a photorealistic Gaussian splatting world of Tartu that scales to the whole city — a BSc thesis can demonstrate the pipeline on a selected district, while an MSc thesis should aim for full city coverage.

Investigating automatic ways to create colliders for Gaussian splats — distinguishing real obstacles from drive-through splats like tree branches and noise — and applying them to a Gaussian splatting world of Tartu for testing autonomous vehicles.

Creating a lane-level navigation graph automatically from recorded decimeter-level (RTK-GNSS) vehicle trajectories, so autonomous vehicles stay in the correct lane and minimize lane changes, and evaluating its quality with established lane-topology metrics.

Creating a taxonomy of scenarios a self-driving vehicle would need to pass to be certified for public streets under EU regulation — especially relevant for end-to-end machine learning stacks — and implementing a subset of these scenarios in simulation (e.g. CARLA ScenarioRunner).

Evaluating existing pre-trained world models and video generation networks for closed-loop testing of end-to-end driving networks, with special attention to rendering different sensors from custom locations and assessing what is feasible beyond cameras, e.g. for lidar.

Adapting Autoware Mini, the lab's autonomous driving software stack, to run on the lab's small-scale test platform based on an electric children's car — finding or writing sensor drivers (lidar, GNSS, camera), connecting them to Autoware Mini and updating the control module, so the vehicle can navigate a small toy city in the Delta courtyard.

Replicating the DrivingBench experiment on the lab's Lexus: setting up a cone course and testing which current frontier vision-language models can drive through it, with repeated independent trials per model and a safety protocol covering safety-driver procedures and model refusals.