Visual Reward for Autonomous Driving
Alho, Lauri (2019)
Alho, Lauri
2019
Tietotekniikka
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2019-05-22
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tty-201905141627
https://urn.fi/URN:NBN:fi:tty-201905141627
Tiivistelmä
Artificial Intelligence (AI) is seen to show wide adaptation possibilities in many fields, and therefore it is used to solve more and more complex problems. One subfield of it is reinforcement learning, which tries to learn a robot to solve a specified task with a given reward function. The reward function is used to tell the robot, how valuable different actions are in different states.
Defining a reward function for a robot in open spaces can be difficult, and one example of this is teaching a robot to drive a car. In these situations, imitation learning and Inverse Reinforcement Learning (IRL) can offer a solution by turning the problem upside down by creating the reward function from expert demonstrations. These can contain any kind of data that the robot uses to learn the correct policy for the task.
This research studies the possibility to use a visual reward for autonomous driving. Driving simulator Carla is used for creating the training data and running the experiments. Expert demonstrations contain driving videos and control data, and latest research results [1] are used for decreasing the required training data to only a dozen of expert demonstrations.
The experiments showed that a visual reward can be used for autonomous driving, when the task is simple. More research should be done for finding working parameters for longer tasks.
Defining a reward function for a robot in open spaces can be difficult, and one example of this is teaching a robot to drive a car. In these situations, imitation learning and Inverse Reinforcement Learning (IRL) can offer a solution by turning the problem upside down by creating the reward function from expert demonstrations. These can contain any kind of data that the robot uses to learn the correct policy for the task.
This research studies the possibility to use a visual reward for autonomous driving. Driving simulator Carla is used for creating the training data and running the experiments. Expert demonstrations contain driving videos and control data, and latest research results [1] are used for decreasing the required training data to only a dozen of expert demonstrations.
The experiments showed that a visual reward can be used for autonomous driving, when the task is simple. More research should be done for finding working parameters for longer tasks.
Kokoelmat
- Kandidaatintutkielmat [8800]