AI-Driven Sport Technique: Pole Vaulting With Proximal Policy Optimization

is a tedious, injury-prone, and personal process for athletes to learn sports. In response, this thesis found a method to use proximal policy optimization (PPO) instead of athletes to create body techniques for sports. The method was applied to a customized Unity ML-Agent that learned to pole vault from scratch without information about how to perform the sport which demonstrated the possibilities of learning without human input, discovered a novel technique, and determined the requirements for developing sport technique with PPO. The reward system encouraged higher body position and clearing the crossbar, while penalizing failures such as hitting the bar or landing early. The agent trained in a competition-style environment, where the bar height increased after a successful attempt and reset to a starting height after three failed attempts. The training resulted in a consistent technique for clearing heights up to 6.30 meters and a maximum clearance of 6.76 meters with a new side swinging technique. Repetitive training at lower heights allowed the model to consistently adapt to higher heights as they were gradually introduced. This research demonstrates the potential of PPO to learn complex sports techniques, and it offers an approach to reward systems by controlling or withholding the level of difficulty at some points and adding repetition.