T↗TONY LABSEXPERIMENTAL SYSTEMS
CYBERPOWER / PROJECT 02
BROWSER ENGINE CPU ONLY
AUTONOMOUS COMBAT EXPERIMENT

AI BATTLE ARENA.

Thirty-two robots. One arena. Intelligence put to the test.

NEUROEVOLUTIONVOL. 001 ↗
● LIVETHE FOUNDRY32-ROBOT FREE-FOR-ALL
00:0032 ALIVE
FDR—01SEED 000031
LIVE SIMULATION
TRAINED NEURAL PROGRAMMED UNTRAINEDClick a robot to inspect its combat telemetry.
CURRENT CHAMPIONLoading artifact…Fixed-topology neural network
GENERATION—32 robots per population
INDEPENDENT RUNS—Separate random seeds
HELD-OUT WIN RATE—Awaiting benchmark results
EVIDENCE, NOT ASSUMPTIONS

The experiment log

Experiment notes ↗
Training artifacts are being prepared. Live simulation is available.
PILOT BRIEFING

How to play

Watch Tournament: choose the robot count, arena, seed and roster, then start. Click a robot to inspect it; Follow, zoom, Sensors and the speed control change the spectator view.

Battle Champion: click or tap the arena to focus it. Use W A S D to move, left/right arrows to aim, and Space to fire. Mouse aim and hold-to-fire are optional. P pauses; R starts a rematch. The pulse cannon deals 20 damage, holds 9 rounds, and reloads automatically.

Train Robots: run a bounded CPU experiment in your browser, inspect the fresh evaluation, then use or export the new brain. Import compatible brain JSON to play it. Higher fitness alone does not prove better combat.

Saved Replays: watch the actual early, champion and final-generation recordings, scrub the timeline, or verify the original simulation. Same-runtime verification can be exact; portable verification separately reports tiny browser numeric drift.

Your device saves: custom brains and replays stay in this browser on this device. There are no accounts or cross-device sync. Export JSON to keep a portable copy; clearing browser storage removes local saves. After the page loads, simulation and learning run on your CPU.

About the experiment and its limits

The shipped brains use a fixed 22 → 8 → 4 neural network with 220 evolved weights. Three independent seeded runs trained populations of 32 for 64 generations. Policies receive current permitted sensors; cover hides opponents, and no future positions or scripted victories are provided.

The displayed benchmark uses fresh fixtures and opponents after validation selection. Across 192 bouts, trained checkpoints scored 6 wins, 149 losses and 37 draws; paired untrained controls scored 1 win, 178 losses and 13 draws. The selected champion scored 5 wins in 64 bouts. Improvements were limited to these experiments; simple programmed opponents remain stronger. Timeout is always a draw, and survival alone never earns a victory.

The champion replay is a real selected showcase, not a typical-outcome claim. Human judgments of game feel, readability and difficulty remain separate from automated verification.