Master Thesis on “High-Level Decision Making in Self-Driving Cars using Deep Reinforcement Learning”

Crossing negotiation at an unsignalized intersection

Decision-making is one of the most important modules in a self-driving car system and navigating an autonomous car in a dynamic and multi-agent urban scenario needs to understand the scene and the interaction between agents. One way to do the scene understanding is to use Graph Neural Networks (GNNs) to learn the interaction between agents and create an embedded vector of the scene for the decision-making module. Then we can use the embedded vector as input to a Reinforcement Learning (RL) algorithm and generate a high-level command for the control module.

Keywords: robotics, autonomous driving, decision-making, reinforcement learning, graph neural network.

Supervisor: Prof. Ville Kyrki (
Advisor: Eshagh Kargar (


  • Review of the relevant state-of-the-art literature on using graph Neural Networks in scene embedding and learning interaction between agents in multi-agent systems.
  • Design some scenarios like unsignalized intersection and roundabout in CARLA simulator with a variable number of cars and pedestrians for training the reinforcement learning agent.
  • Implementing the scene embedding and design the pipeline to train the RL agent.

Practical Information

Pre-requisites: Deep Learning, Reinforcement Learning, Autonomous Driving.
Tools: Python, PyTorch.
Start: Winter semester


  • Agarwal, A., Kumar, S., & Sycara, K. (2019). Learning transferable cooperative behavior in multi-agent teams. arXiv preprint arXiv:1906.01202.
  • Diehl, F., Brunner, T., Le, M. T., & Knoll, A. (2019, June). Graph neural networks for modelling traffic participant interaction. In 2019 IEEE Intelligent Vehicles Symposium (IV) (pp. 695-701). IEEE.