Training an AI to play Resident Evil Requiem using Behavior Cloning + HG-DAgge [P]

A developer has successfully trained an AI agent to master a challenging segment of the game Resident Evil Requiem, demonstrating a practical application of advanced imitation learning techniques. Instead of using pure reinforcement learning, the project employed a hybrid strategy: Behavior Cloning (BC) provided the initial policy by learning from human gameplay demonstrations, while the HG-DAgger algorithm was used to iteratively refine the agent's performance and correct the compounding errors common in BC when the agent deviates from the training data.

The environment was built from gameplay capture, with controller inputs mapped to a discretized action space and observations taken directly from preprocessed video frames. Key challenges included early training instability and the critical need for precise action-to-screen synchronization, where even minor timing mismatches could derail learning. After training, the agent achieved consistent navigation through the semi-linear escape sequence, demonstrated real-time reaction to enemies, and showed a degree of recovery from unexpected deviations. The code for the project, which remains specialized to this specific game segment, is publicly available on GitHub, offering a valuable case study for applying BC and DAgger variants to complex, time-sensitive virtual environments.