SAE - GFR Reinforcement Learning for Trajectory Planning for Autonomous Racing

Reinforcement Learning is the process of training a policy to maximize the reward from an environment via large scale trial and error. Previous years iterated on the reward function. The first part of the project will be to introduce random states for acceleration and steering angle into the environment in order to increase robustness. The second part of this project will be improving the training architecture by implementing the distributed model and specifically training on GPUS when possible. This project will require use of the High Performance Cluster available to OSU students in order to speed up training as well.

Objectives


  • Fall
    • Background research of previous projects and relevant papers completed.
    • Github and development environment set up.
    • ROS tutorial and assignment completed.
    • Design changes conceptualized.
    • Practice for participation in rules quiz.
    • Documentation for any changes made during Fall.
  • Winter
    • Final documentation for code completed - details on wiki as well.
    • Testing and benchmarking completed and documented.
    • UML Diagram detailing system.
    • Final edits to the code by May.

Motivations


Reinforcement learning is an opportunity to simplify our controls and perform cutting edge work with an autonomous vehicle. Continuing the research from previous years and working on a budding repository is useful as well.

Qualifications


Minimum Qualifications:
None Listed

Preferred Qualifications:
None Listed


Details


Project Partner:

Marcus Wheeler

NDA/IPA:

No Agreement Required

Number Groups:

1

Project Status:

Accepting Applicants

Website:
https://www.global-formula-racing.com/en/
Keywords:
Machine Learning (ML)racingAutonomousReinforcement LearningTrajectory Planning
Card Image Capstone