Visual Place Recognition Pre-Training for End-to-End Trained Autonomous Driving Agent

End-to-end autonomous driving often relies on the concept of learning to imitate from expert demonstrations. Since those demonstrations cannot cover all possible variations in data, there always are situations where the trained agents encounter unseen conditions, which results in a shift in the data distribution. One of the most common causes of this shift is changes in weather and lighting conditions. In this study, we suggest using a pre-training based on the visual place recognition (VPR) method, in order to mitigate this effect. We compare the corresponding navigation agent to a baseline agent which relies on the commonly used ImageNet pre-training by evaluating as per the Leaderboard driving benchmark in CARLA environment. According to our experiments, pre-training on the VPR task shows higher resistance to unseen weather conditions. The findings calculated in our study are evaluated over multiple seeds to show statistical consistency. The accompanying open-source code repository can be accessed via
https://github.com/Shubhamcl/vpr_pretrained_agent/
.