Obstacle Tower Challenge

Arcade Learning Environment (ALE) is one of the most sought-after game environments used in reinforcement learning research so far. Such games are two-dimensional, have low-pixel graphics resolutions, operate in a fixed layout, have limited branching factors and do not require agents to be trained for generalization.

​

The Obstacle Tower Challenge is a high-fidelity game that operates in 3D, generates new visuals and themed layouts with each level and game using “Procedural Content Generation”, has a high branching factor, and requires agents to learn to solve sub-tasks. Although this game is more advanced than before, this adds those many challenges to developing agents that can match human-level performance. RL researchers are focusing on developing autonomous agents that can perceive more information, strategize as well as control low-level actions to achieve high scores in this game.

​

Our goal in this project was to develop and train an AI agent using state-of-the-art reinforcement learning techniques that can successfully maneuver the OTC game and solve an above-average number of levels. This project also helped us gain hands-on knowledge about generalization in AI and its challenges.