Project Pleiads

Project Pleiads is the first Arcade.it AI Lab experiment. We are building a reproducible system that can observe, play and analyse the 1981 arcade game Pleiads / Pleiades, by Tehkan, using local AI, computer vision and automated search.

Best verified Search

5,750

points · zero model calls

Perception baseline

90%

9B / history3 / nearest4

Opposite-side error

1.5%

on the same 200 visible-player cases

Current experiment

E006

E006 — Diversity Search

5,464 completed registry records in the last full dataset checkpoint at this checkpoint, including infrastructure smoke tests. E005: 2,002. E006: 3,102.

Full registry counts last synchronized: 2026-10-05T08:28:16.420013+00:00.

Watch the latest verified record

Latest verified milestone / run evidence
VERIFIED FROM INGESTED ARTIFACTSAUTOMATED POLICY SEARCHMODEL CALLS: 0REAL-TIMEREPLAY VERIFIED
Run ID
pleiads-20261005-131958-e6ceb4
Experiment / controller
E006 / native5
Score
5,750 points
Survival
216.760 s (emulated game time)
Real-time / replay
Confirmed / exact recorded-input replay
Screenshot matches
2,239 / 2,239 captured samples
Original manifest SHA-256
9df9cf29b00e0b2ad6f2d209b462bff7f27495a0c01c38b14e83a7c013ec3943

All captured samples were checked, at one capture every six emulated frames. Screenshot bytes and game software are excluded. CSV preserves the original TSV action order. This is a laboratory milestone, not a world-record claim.

Validated checkpoint: 2026-10-05. Data last synchronized: .
Latest record verified from reviewed published run evidence. Registry counts retain the last full dataset checkpoint; record publication does not refresh the complete dataset. This is a snapshot, not a live counter.

Registry count checkpoint: 2026-10-05T08:28:16.420013+00:00. The latest reviewed record and its evidence update separately from full dataset totals.

Frames, models, skills and game inputs

MAME
captured framesPerception
CV / optional QwenController
legal joystick + fireMAME
next game frames
QWEN / MODEL BRANCH

Frames → Qwen observation → strategy / skill selection → controller

Perception benchmark: E002. Policy selection experiment: E004. These results carry their own model and latency evidence.

SEARCH / ZERO MODEL CALLS

Policy Search → timed skills / native plan → controller → MAME

E005 and E006 milestone runs execute the selected plan directly, with no Qwen inference during gameplay.

Original laboratory architecture. Manifests, ordered actions and hashes leave the research machine through quarantine, validation, review and atomic publication.

Four distinct research branches

01 · Perception / Qwen
Local multimodal models observe game frames. E002 measures player localization against independent sprite-based ground truth. A perception result is not a gameplay score. Read E002.
02 · Controller / computer vision
Fast controllers turn perception and policies into valid joystick and fire inputs. Observation, decision timing and input execution are logged separately. Read E003.
03 · Automated Policy Search
Search explores movement and firing policies without asking an LLM to decide every frame. There are zero model calls during a Search run. Accelerated discovery is followed by real-time confirmation and replay checks. Read E005.
04 · Strategy / higher-level models
Models can classify situations, analyse failures, compare policies and select skills. E004 tested policy selection; that role did not automatically outperform deterministic Search. Read E004.

What the experiment can establish

We distinguish an observed score, a real-time confirmation and a reproduced input sequence. Evidence labels describe the strength and origin of a claim. We make no claim that the game has never been completed, and no world-record claim.

The laboratory uses MAME and game software obtained separately. We distribute our experiments, not the game. Read the ROM and media policy and reproducibility guide.

A short history of the game

MAME’s maintained driver identifies Pleiads (1981), Tehkan, with a licensed Centuri version. Its input definitions specify horizontal left/right movement and a firing button. The laboratory preserves the emulator version and DIP settings in each run manifest. Source: MAME’s Phoenix/Pleiads driver.

In our observed runs, the player moves along the lower edge, fires upward and encounters changing enemy formations. Those transitions motivate timed skills and visual controllers. Their exact phase identities, round counter and any late-game ending remain unverified laboratory questions; we do not infer them from a score alone.