Match
Blue = the agent · Red = you or a scripted opponent. First death loses.
Blue0
Red0
Offline maze tank duel with bouncing shells, walls, power-ups and a shrinking poison ring — play against, or watch, a tank agent trained by imitation + self-play under random spawn.
★ v1.5 neural agent — mean 0.59 win-rate vs the laika family (laika · easy_laika · stationary · pro)
Set up any match in the bar at the top of the page — pick the 地图 Map (1v1 duel, random or fixed spawn)
and who drives each tank: 蓝方 Blue / 红方 Red can each be 玩家 (Player),
a ★ 智能体 agent, or a hand-coded 脚本 script. To play yourself, set a side to 玩家 and click the arena to
capture keys (WASD+Space = blue · Arrows+Enter = red). K cycles the camera, [/] change speed.
Blue = the agent · Red = you or a scripted opponent. First death loses.
The agent (★ v1.5 neural) is a 105-d, identity-blind policy (it never sees the opponent's name); the other options are hand-coded script bots.
Honest note: it wins ~59% overall here, but the evasive laika remains a hard ~30% — a documented
"fire-timing" wall. Full write-up in docs/TRAINING_HISTORY_en.md.