Definition
Plain language
A classic box-pushing puzzle game used as a planning benchmark for AI.
As stated in the literature
A grid-based puzzle environment requiring spatial planning, widely used as a benchmark for RL and planning agents.
Why it matters: Irreversible moves make Sokoban an unforgiving test of true planning rather than trial-and-error, which is exactly why it's stuck around as a benchmark.
For example, the agent has to push boxes onto target squares in a warehouse-like grid, but pushing a box into a corner makes the puzzle unsolvable forever.
Heard on the show
“On the Sokoban environment, the lowest-variance quartile gets you about eleven percent task performance.”Episode 010 — When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL