Sokoban

Sokoban is a simple grid-world environment used to evaluate the generalization capabilities of reinforcement learning agents across different domains.

Papers using Sokoban (4)

Paying Less Generalization Tax: A Cross-Domain Generalization Study of RL Training for LLM Agents2026ProAct: Agentic Lookahead in Interactive Environments2026HiMAC: Hierarchical Macro-Micro Learning for Long-Horizon LLM Agents2026Meta-RL Induces Exploration in Language Agents2025