Game AI experiments
Watch three experimental controllers attempt Doom. Choose a player to follow, or compare all three. These are spectator controls; the players make their own moves.
Have a game AI idea? Ask us about your project, or explore our bots and game services.
01 / Reference dynamics
The retained MaleCNS network and original v6 learning rule.
Connecting to Original Doomfly...
Open to connect to this controller's observed activity.
02 / Compiled dynamics
The retained network and learning rule in the compiled ERAIS bank, with adaptive neural cadence.
Connecting to Fracture full connectome...
Open to connect to this controller's observed activity.
03 / Adaptive controller
A native neuronal expert alongside ERAIS visual circuits, adaptive association and persistent memory. Source status reports the active composition.
Connecting to ERAIS cyborg fly...
Open to connect to this controller's observed activity.
Gameplay updates as fresh snapshots arrive. Pause keeps a timestamped still image while the player keeps running. Open an inspection panel to see available activity and visual inputs.
The players use different neural timing. These independent live attempts are observations, not a matched performance or learning benchmark.
The viewer requests at most four gameplay snapshots per second for each visible player. Pixels and telemetry come from the same acquisition. Live frames expire within five seconds. Pause keeps a timestamped still while the player continues; resuming replaces it when a fresh snapshot arrives.
The players are the original fly-network simulation, the complete network running through ERAIS Fracture, and a cyborg combining that native network with an adaptive executive.
Player credit records the engine-confirmed killing blow. World deaths also include infighting and environmental damage. A round is a complete attempt; averages cover completed rounds in the verified tracking window and its latest 100, with the current sample count shown above. Interrupted attempts are excluded. Learning and statistics persist across runs.
Freedoom: Phase 2, normal difficulty, indexed seed schedule starting at 41027. Level transitions retain equipment; game over returns to MAP01. All runs use 35 simulated game tics per second. The reference follows its original neural clock; the native gameplay variant currently uses adaptive neural cadence and can hold a previous interval's counts. Both retain the declared network, but their live timing is not equivalent. Transport rate is not simulation speed.
RSS is currently resident memory. The additional memory figure includes the process’s proportional share of resident and swapped pages, so a swapped-out player does not appear artificially small. These are whole-process measurements, sampled every two seconds.
Explore movement, sampled neural activity and vision in a simulated fly with Fly Lab. Its measurements show how quickly the simulation is running.
From neural circuits to native computation
Fly circuits offer ideas for processing information: selective amplification, inhibition, continuous state and outcome-dependent adaptation. We test the underlying mathematics before adopting a faster implementation.
Color contrast, shape and spatial competition produce visual cues. Weapon outcomes train their credibility: unsuccessful firing can reduce attraction to static marks. The controller sees RGB pixels, without enemy labels or hidden target coordinates.
Directional state, sparse association and homeostatic signals contribute to movement, aiming, interaction and weapon selection. Health and ammunition provide body-state inputs. Navigation learns action values from visible change and subsequent outcomes.
Visual credibility, navigation values and associative memories persist across attempts. Fast sensory traces reset at a new life so the next scene does not inherit the previous death’s credit. Matched frozen-learning runs test whether the updates improve play.
The primary ERAIS cyborg adds a real, complete native MaleCNS neuronal expert to that learned executive. Its source weights are frozen; modeled neuronal state resets for each game episode. Actual descending-neuron counts contribute to an explicitly engineered steering and forward-motion composition. The executive continues its own visual, navigation and associative learning, with its functional component states included in new checkpoints. This is an experimental controller composition, not a claim of donor-action equivalence or improved gameplay.
Original Doomfly runs the full retained MaleCNS v1 network: image luminance drives mapped R1–R6 inputs, RGB proxies drive 811 mapped R8 cells, and neural spikes reach its fixed BCI motor decoder. Nonfatal damage supplies a subsequent PPL101 pulse. Its original baseline-centered plasticity rule updates 4,184 existing KC-to-MBON11 connections; learned and fast neural state persist between attempts. The players share the campaign interface’s interaction and empty-ammo weapon-switch extension.
The Fracture player runs the complete retained v6 network through the standalone ERAIS native bank, with 0.1 ms executed neural steps and the centered plasticity rule. Its gameplay adapter uses adaptive neural cadence. The cyborg also uses multiple time scales: its complete neuronal expert advances in 20 ms intervals under the executive’s cognitive cadence, while its adaptive K selects functional associations. Every invoked native interval retains the full neuronal graph. Neither adaptive schedule establishes matched-game-time donor equivalence. The observation panel identifies frozen expert state and executive learning separately.
Vision panels prefer the last image actually consumed by the neural bank, so a held neural interval is not paired with a newer gameplay image. The reconstruction is partial receptor-state luminance, not whole-brain scene decoding. Its reprojection error measures receptor consistency; it is not image accuracy. Source connectivity, inferred display mappings and chosen physiology have distinct evidence boundaries. Live source metadata remains authoritative if a fallback publishes a different composition.
The fly HUD portraits mirror pleasure, aversion, health and attack state. Shared campaign body inputs track ammunition expenditure and resource acquisition; a conservative stuck-loop supervisor retains learned experience across restarts.
Built with ViZDoom and packaged Freedoom assets. NeuroForge is not affiliated with id Software. The research goal is one adaptive architecture and model bank across these experiences; shared-stack integration is ongoing.