Cortex: Compact Quake Policy with Frozen Visual Features

The research paper titled 'Cortex: Compact Behavior Cloning for Quake with Frozen Visual Features' presents a behavioral-cloning policy designed for the first-person shooter Quake. Utilizing 10.98 million trainable parameters, this policy employs a six-layer transformer built on a frozen DINOv3 encoder. It is trained using the Quake subset from the public Pixels2Play corpus, which includes 6,849 recordings totaling approximately 474.7 hours, represented by 17.09 million cached decision frames detailing keyboard and mouse actions. Each training epoch samples 517,048 four-frame windows and requires 3.3 minutes of policy-head optimization on a single RTX 5080, excluding the initial feature extraction. Although Cortex does not finish the level, it successfully reaches the opening door, button room, and gate descent in every episode, with 19 of 20 episodes in each batch achieving at least one kill. The research investigates the potential of a straightforward behavioral-cloning policy prior to incorporating reinforcement learning or explicit memory.