Carlman

Level1/6

Score0

adding mass removing mass target direction

Maze-chase game played inside Lenia, a continuous cellular automaton. Carlman is steered by CARL (Cvjetko et al. 2026), a reinforcement learning agent that keeps him on course by adding or removing a little mass at a time. The Carlman, the ghosts and the dots are all solitons: self-organizing patterns that move and hold together by the automaton's own rules. Eat every dot to clear the level and move on to more ghosts, grab a power pellet to make the ghosts edible for a while, and keep away from them otherwise: a ghost's touch dissolves Carlman. Use the arrow keys, or click or tap the maze, to choose which way Carlman goes.

Settings

? mass 0 steps 0 CARL decision – measured speed –

action cost ? 5.00 action threshold ? 0.15 game speed ? 60/s

model

Model stats

action frequency – held back – exploded 0 died 0

action frequency: how often CARL adds or removes mass.
held back: how often the action threshold stopped CARL from acting.
exploded: times Carlman's mass ran away.
died: lives lost, ghosts included.

These reset when you restart the game or pick another model.

Performance. The game runs entirely on your device. Models with a smaller Lenia kernel radius r or fewer parameters may run faster. If the game stutters, try one of those, or lower the game speed under Settings.

soliton ?

★ is the most stable. The other two may be unstable, and may need a different action cost or model.