Background
Autonomous driving has advanced rapidly in recent years, especially for passenger vehicles.
However, transferring these developments to trucks introduces added complexity due to their
size, limited maneuverability, and stricter safety demands. A major challenge is the lack of
diverse, truck-specific training data. Most existing datasets are biased toward passenger cars
and simple, straight-driving scenarios, limiting their usefulness.
The primary goal of this project is to explore methods to adapt and repurpose car-based
datasets for truck motion prediction and planning. Specifically, the project aims to develop
strategies that identify diverse behavioral patterns and account for truck-specific constraints,
enabling better generalization and safer planning.

Methods and Milestones
To bridge the gap between car and truck data, the project will explore a multi-stage approach.

1. Scenario Selection and Tagging:
Using Vision-Language Models (VLMs), Variational Autoencoders (VAEs), or rule-based
systems, students will identify and label scenarios in car datasets that are potentially
transferable to truck contexts. This preprocessing step will filter out irrelevant or
infeasible driving scenes and make the dataset more balanced.
2. Trajectory Adaptation via Optimization:
Feasible car trajectories will be transformed into truck-compatible trajectories using
trajectory optimization techniques, such as Model Predictive Control (MPC) to
incorporate truck-specific dynamics.
3. Simulation and Model Evaluation (optional):
Simulators will be used or enhanced with truck-specific constraints to allow for more
realistic evaluations of the trained models.


In the different project stages, the students will evaluate a baseline model trained on car data
against the same model trained on a curated dataset, a dataset with optimized truck-
trajectories and the combination of both approaches.

Available Data
To accelerate the initial learning curve, students will work with large open-source datasets.
The nuPlan dataset includes over 13,000 hours of driving logs (accounting for 2TB of data),
while Argoverse provides 250,000 scenarios across 2,000 kilometers, adding geographic and
behavioral variety. For evaluation on real truck behavior, a smaller proprietary dataset
from Traton, containing around 30 hours of truck driving data, will be used to fine-tune and
validate the models.

Requirements

  *   Strong knowledge of machine learning.
  *   Experience with Python and deep learning frameworks.
  *   Familiarity with autonomous driving is a plus.

Apply to this project here