This repository provides a Sokoban environment for OpenAI Gym, designed for reinforcement learning research. It addresses the challenge of irreversible mistakes in Sokoban puzzles by implementing a novel, solvable level generation algorithm based on reverse play and a heuristic scoring system.

How It Works

The environment generates random, solvable Sokoban levels using a three-phase process: topology generation via random walks, element placement (player, boxes, targets), and a crucial reverse-play phase using Depth First Search to ensure solvability and assign a difficulty score. This approach allows for training RL agents on diverse, non-overfitting scenarios.

Quick Start & Requirements

• Install via pip: pip install gym-sokoban
• Requires Python.
• Official docs: https://github.com/mpSchrader/gym-sokoban

Highlighted Details

• Implements 9 actions (Move/Push in 4 directions, No-Op).
• Offers multiple rendering modes: rgb_array, human, tiny_rgb_array, tiny_human.
• Includes various room configurations (e.g., Sokoban-v0 to Sokoban-huge-v0) with different sizes and box counts.
• Supports variations like Fixed Targets, Multiple Player, Push&Pull, and Boxoban (DeepMind puzzles).

Maintenance & Community

• Project initiated by Max-Philipp B. Schrader.
• Open for contributions via issues or pull requests.

Licensing & Compatibility

• License not explicitly stated in the README.

Limitations & Caveats

• Larger room configurations may take significant time to generate, especially on laptops.
• The README does not specify the exact license, which could impact commercial use.