Learned driving

Creating modular autonomy software can be complex and prone to failure. An alternative approach, called "learned driving", aims to simplify the process by using a single neural network instead of a complex stack of modules.
However, this approach also has its drawbacks, as neural networks can be unpredictable and may fail without warning. Our research aims to improve the reliability and robustness of machine learning-based self-driving techniques.

So far, we have accomplished the following:

  • Collected a dataset from WRC Rally Estonia tracks across all four seasons.
  • Successfully demonstrated learned autonomous driving using this dataset with both camera and lidar sensor inputs.
  • Created a Vista simulation of these rally tracks, which greatly simplifies the evaluation of trained models.

Our long-term plan is to begin with a basic task of road following, then progress to lane following on highways, and eventually reach urban driving scenarios.

Tambet Matiisen

Tech Lead, Autonomy Software and Learned Driving Team Lead

tambet.matiisen@ut.ee
Group channels
Research papers
LiDAR-as-Camera for End-to-End Driving Controlling Steering with Energy-Based Models A Survey of End-to-End Driving: Architectures and Training Methods
Theses
Physical A*: Graph-Based Search Algorithm for Robot Navigation On-the-Go Asymmetric Deep Multi-Task Learning Energy-Based Models for End-to-End Autonomous Driving Comparing Output Modalities in End-to-End Driving Navigation Application Input Rendering for a Self-driving Car’s Neural Network Using LiDAR as Camera for End-to-End Driving