09 / Reinforcement learning
DOOMORPH
Watch a recurrent combat policy across seven held-out scenarios.
Recorded evaluation · checkpoint at 1,054,720 decisions
09 / Held-out evaluation recordings
Watch the policy under pressure.
Select a scenario to watch its saved evaluation and read the measured result. The recording overlays actions, value estimates, and policy gates.
Selected scenario
Line defense
20.2kills per episode across five held-out episodes
Stationary combat is this checkpoint's strongest behavior.
Loading scenario measurements…
Survival across tasks
The project
Recurrent visual and audio control
DOOMORPH uses game pixels, stereo audio, previous action, and recurrent state in a PPO policy. The selected checkpoint was trained for 1,054,720 decisions. This suite evaluates seven scenarios with five held-out episodes each.
Observations & lessons
What the suite taught us
The policy learns stationary combat more readily than navigation or survival. It averages 20.2 kills in line defense, but reaches no exits in the navigation tasks. Strong left-turn and continuous-fire habits remain visible in the recordings. Behavioral gates expose those habits even when reward improves.