In the project, we plan to develop an AI agent to parkour through a map of obstacles. Inspired by minecraft parkour maps, a website that creates parkour maps for human players to challenge themselves, we plan to create our own version of parkour maps that are suitable for our agent to practice on. The maps will be 3D maps varying in difficulty with obstacles, such as walls and ladders, that require the agent to jump over while finding the fastest route to reach goal point(s). Our agent will take in the information that is the parkour map and produce the series of actions to run through it as efficiently as possible.
Since the bot will be attempting to beat the maps within a specified time limit, reinforcement learning using neural networking best fits this goal.
The main goal is for the AI to be able to reach a certain goal point(s) as quickly as possible while avoiding any obstacles or pits. As the AI goes through more trials and runs into obstacles, it should be able to learn from its mistakes and explore different paths to decide the fastest successful route to the end goal or midpoints. In time, it should be able to consistently choose the best path to be equal to or better than the average human player in reaching the goal.
The AI will be analyzed/verified by checking the time of the route as well as whether it misses any midpoints in applicable courses as midpoints are increased in number. The “moonshot case” would be for the AI to be able to learn from completely new and complicated maps found online, such as on minecraftmaps.com and be able to find the fastest solution to these maps created by separate users.
October 23, 2pm PST