GraphSCENE:
On-Demand Critical Scenario Generation for Autonomous Vehicles in Simulation
Efimia Panagiotaki, Georgi Pramatarov, Lars Kunze, Daniele De Martini
Oxford Robotics Institute, Department of Engineering Science
University of Oxford
IEEE IROS 2025
A tool to automatically generate diverse, safety-critical traffic scenarios on-demand in simulation for autonomous vehicles testing.
Overview
Reliable AV deployment requires testing in realistic, safety-critical, and diverse scenarios, yet manually designing such environments is costly and inefficient.
GraphSCENE solves this by automatically generating dynamic traffic scenarios tailored to user preferences. Based on temporal graph neural networks (GNN) and grounded in real-world traffic priors, GraphSCENE produces scenarios that are rendered directly in simulation.
GraphSCENE solves this by automatically generating dynamic traffic scenarios tailored to user preferences. Based on temporal graph neural networks (GNN) and grounded in real-world traffic priors, GraphSCENE produces scenarios that are rendered directly in simulation.
GraphSCENE reduces the overhead of manual scenario creation, making AV testing more scalable, realistic, and effective.
🚦 Realistic traffic dynamics - scenarios follow real-world behavioural constraints.
⚡ Scalable & on-demand - generate multiple scenarios fast, without manual design overhead.
🎯 Tailored to user preferences - users define the types of situations they want to test.
🧩 Seamless simulation integration - outputs are rendered in CARLA Sim.
🛡️ Safety-critical coverage - stress-test AVs under diverse and challenging conditions.
⚡ Scalable & on-demand - generate multiple scenarios fast, without manual design overhead.
🎯 Tailored to user preferences - users define the types of situations they want to test.
🧩 Seamless simulation integration - outputs are rendered in CARLA Sim.
🛡️ Safety-critical coverage - stress-test AVs under diverse and challenging conditions.
Methodology
1. Scene Graph Extraction: Temporal scene graphs are extracted from real-world datasets, representing dynamic agents, static structures, and their relations, to create a database of seed scene graphs.
2. Scenario Specification: Users define the type of scenario to test (criticality, AV action, agents).
3. Graph Completion: A temporal graph neural network (GNN) predicts ego-vehicle interactions and relations over time, constrained by traffic priors and semantic rules.
4. Validity Constraints Predictions are filtered through a domain ontology and real-world priors to ensure realistic, rule-consistent interactions.
5. Simulation Integration The final scene graph is converted into CARLA Sim, producing scalable, safety-critical scenarios for AV validation
2. Scenario Specification: Users define the type of scenario to test (criticality, AV action, agents).
3. Graph Completion: A temporal graph neural network (GNN) predicts ego-vehicle interactions and relations over time, constrained by traffic priors and semantic rules.
4. Validity Constraints Predictions are filtered through a domain ontology and real-world priors to ensure realistic, rule-consistent interactions.
5. Simulation Integration The final scene graph is converted into CARLA Sim, producing scalable, safety-critical scenarios for AV validation
Results
We compare our method against a few-shot LLM, a state-of-the-art link predictor, and a graph perturbation approach, assessing link prediction accuracy, contextual correctness, and scenario generation consistency and plausibility. GraphSCENE consistently outperforms all baselines, particularly in generating semantically valid ego-agent interactions and maintaining temporal coherence across scenes. Outputs are directly integrated into CARLA simulations, confirming both fidelity and utility for AV validation pipelines.
BibTex
@misc{panagiotaki2025graphsceneondemandcriticalscenario,
title={GraphSCENE: On-Demand Critical Scenario Generation for Autonomous Vehicles in Simulation},
author={Efimia Panagiotaki and Georgi Pramatarov and Lars Kunze and Daniele De Martini},
year={2025},
eprint={2410.13514},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2410.13514},
}