Seminarai, kursai

2025 m. kovo 31 d., 13 val.

Vilnius, Akademijos g. 4, 203 kab.
Nuotoliniu būdu „MS Teams“ aplinkoje (https://bit.ly/DMSTI_2025-03-31)

Shubham Juneja

(disertacijos pristatymas,
vadovas: prof. dr. Virginijus Marcinkevičius)

„Investigation of Pre-training in Imitation Learning Based Autonomous Driving“

Anotacija: Autonomous vehicles promise transformative changes in transportation, with SLAM-based methods enabling map-based navigation and learning-based approaches leveraging neural networks for data-driven decisions. While SLAM provides map-based navigation, learning-based methods leverage neural networks for data-driven decisions. This study centers on imitation learning within the learning-based paradigm, addressing its limitation, co-variate shifts. The aim is to develop autonomous navigation systems using deep learning and imitation learning, emphasizing pre-training techniques. This research starts off with reviewing state-of-the-art imitation learning methods and pointing out how pre-training in autonomous driving is under-explored. Majority approaches in this area of research choose visual encoders pre-trained on the task of ImageNet classification, rather than searching for better alternative approaches. Therefore, the study proposes application of pre-training methods novel to the task of end-to-end autonomous driving. It then evaluates these methods against baseline approaches to demonstrate enhanced performance.