Effect of delays/lag and fighting it in self-driving neural networks

Organization
Autonomous Driving Lab
Abstract
For camera-based self-driving neural network models, we collect a dataset of image and driving command pairs. This dataset connects a camera image with how the human driver steered the car at that moment. We then train a network that would take as input the camera feed and would output the driving command (predicting what would a human do here?).

However, the networks take time to compute their answers and the command reaches the wheels with a delay. Also, the motors turning the wheels have intrinsic delays. This means that even if the model can perfectly predict what a human would do in the situation, the answer is found too late and the car will not stay on track.

In this project:
-You evaluate the amount of delay, T milliseconds, the network computations infer
-You train models that predict not the steering command captured at the same time as the image, but T milliseconds after that. So the network would incorporate our knowledge about its delays.
-Compare prediction accuracy - is predicting future command even possible based on the present image?
-Compare driving ability - will the cars stay on the road better?

This project will be mainly done with toy cars. Once we prove the concept that predicting future commands is useful, a model can be trained also on real data for the Lexus. Performance on the Lexus will be evaluated qualitatively (does it feel safer, smoother?).
This may be publishable work.
Graduation Theses defence year
2021-2022
Supervisor
Ardi Tampuu
Spoken language (s)
Estonian, English
Requirements for candidates
Level
Masters
Keywords
#self-driving, AI, neural networks, deep learning

Application of contact

 
Name
Ardi Tampuu
Phone
E-mail
ardi.tampuu@ut.ee