Academic
Criticality-Aware Motion Planning
Bachelor thesis on trajectory scoring for autonomous-vehicle motion planning using Transformer models and heterogeneous GNN context.
Key results
The model was evaluated as a trajectory-ranking surrogate for the Reactive Planner. The main question was whether it could preserve the planner-induced candidate ordering.
Overview
This bachelor thesis explored how learned trajectory scoring can support autonomous-vehicle motion planning.
Sampling-based planners generate many feasible trajectory candidates and then rank them using a cost function. The thesis focuses on learning a trajectory evaluation model that predicts planner costs while using scenario-level criticality context.
Problem
In each planning cycle, the Reactive Planner generates candidate trajectories, filters infeasible options, ranks the remaining candidates by cost, and checks them for collisions.
The ranking step is important because the selected trajectory depends directly on the ordering induced by the cost function. When many candidates are generated, cost evaluation can become expensive, especially if additional scenario context or criticality information should be considered.
Approach
The proposed model combines three parts:
- candidate trajectories generated by the Reactive Planner
- scenario context from a heterogeneous GNN
- a Transformer-based trajectory scoring model
The GNN provides a scenario-level criticality probability and scenario embedding. The Transformer uses this context together with trajectory features to predict per-trajectory log-costs and an auxiliary criticality score.
Dataset and evaluation setup
The model was evaluated on CommonRoad scenarios and planning cycles generated by the Reactive Planner.
| Item | Value |
|---|---|
| Scenario dataset | CommonRoad-based scenarios |
| Evaluated ranking cycles | 2,604 |
| Main task | Ranking candidate trajectories |
| Main model | Transformer-based trajectory scorer |
| Context model | Heterogeneous GNN scenario encoder |
| Main ranking metric | NDCG@3 |
| Final thesis grade | 1.0 |
Ranking evaluation
The strongest results are shown in the metric cards above.
The model reached strong top-ranked candidate performance, including P@1 = 0.79, P@3 = 0.90, and NDCG@3 = 0.98. It selected the exact best-cost candidate in 77.8% of evaluated cycles, and a candidate within 5% of the best cost in 92.7% of cycles.
Planner-level evaluation
The learned model was also evaluated inside the Reactive Planner by replacing the baseline ranking cost with the predicted cost.
On the additional evaluation dataset, the learned model matched the baseline planner outcome with 359 passed and 57 failed scenarios.
What I learned
This project connected autonomous systems, graph neural networks, Transformer-based sequence modeling, ranking metrics, and planner-level evaluation. It also required turning a research idea into a complete implementation and evaluation pipeline.