AUTONOMOUS DRIVING: LEARNING FROM INTERVENTIONS
Keywords
Loading...
Authors
Issue Date
2021-10-01
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Autonomous driving systems have not yet reached full self-driving capabilities. Safety
driver oversight is required to recognize challenging situations and detect deviations.
Each time the safety driver intervenes, a failure mode of the autonomous driving system
is signaled. In this work, methods to improve deep neural network models based on
data springing from safety driver interventions are evaluated. Self-driving models are
usually trained through regular imitation learning on expert examples. To incorporate
interventions in such training, a novel negative learning concept is proposed that is similar
to reinforcement learning. The experiments are performed in Carla, a realistic driving
simulation. Offline learning from interventions is found to improve model performance,
though care has to be taken as overfitting is likely when many interventions are similar.
Negative neural network learning is found to have a small, but promising, effect.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen