DeathBot: Neural network for 20G Tetris play

DeathBot/TAPNet is a neural network trained to survive and achieve Tetrises in 20G, designed for the Tetris: The Grand Master series of games (specifically TGM2's T.A. Death mode). It is a Pytorch neural net (430k parameters) built on a Rust simulation of TGM game mechanics, with a bridge to the BizHawk emulator for live play.

The 1-ply champion (snapshot12) is pure greedy play with no Tetris-scoring ability. It can survive for ~2,150 pieces on average, meaning it is good enough to achieve GM in most Death games. The model's abilities can be expanded to practically endless survival with 2-ply (search over next piece placements), which can be tuned for "stylish" play (scoring Tetrises). The stylish default setup is cool30lam10tiered_burn6 on snapshot12 (commitment controller + λ10 dense well potential, tiered preserve, burn6), which achieves a Tetris rate around 22-24% of the theoretical maximum. Style is currently achieved only through search and the controller, not the weights (self-distillation for stylish play was unsuccessful).

Layout

• model_weights_snapshot12.pt- final weights ("TAPNet", 430,465 params). Snapshots 1–11 are the training lineage (same arch)
• deathbot_sim/ - Rust sim (src/: board, ARS rotation/kicks, IRS, 20G, BFS search_locks) + Python (python/: model.py, two_ply.py, selfplay.py, trainers, visualizer)
  • deathbot_sim/TRAINING_ARCHIVE.md - condensed training history and learnings (replaces the old unsorted 5000-line PROJECT_STATE.md)
  • deathbot_sim/_diag_out/PANEL_REGISTRY.md - seed-panel discipline record
  • deathbot_sim/_diag_out/ - key promotion/showdown evidence JSONs only. Intermediate checkpoints, banks (.pt/.npz), and logs were removed; regenerate with the drivers in python/
• tgm_bizhawk/ - live-play bridge: tgm_1ply.py (snapshot12 1-ply suggester), tgm_mirror.py (viewer), tgm_player.lua (BizHawk state publisher + input playback). See BIZHAWK_STATE.md + MEMORY_MAP.md. You will need your own copy of BizHawk and a TGM2P ROM.

Quickstart

From deathbot_sim/ with the workspace venv:

• Rebuild sim: <venv>/Scripts/python.exe -m maturin develop --release
• Tests: cargo test and python -m pytest tests/ (needs model_weights_snapshot{7,9,11}.pt at root + _diag_out/state_bank_snapshot9_32games.npz fixture)
• Sweep: python python/sweep_eval.py --help
• Live viz: python python/viz_live.py --two-ply
• Live bridge: start tgm_player.lua in EmuHawk (T.A. Death), then python tgm_mirror.py from tgm_bizhawk/.