lmenv

Environment implementations for GRL games (Tetris and Sokoban). This repository provides self-contained, deterministic game environments suitable for reinforcement learning and language model training.

Overview

lmenv contains implementations of classic puzzle games adapted for use in the GRL (Game Reinforcement Learning) framework. The environments are designed to be:

• Deterministic: Reproducible game states with seed-based generation
• Text-based: Human-readable game representations for language model interaction
• Self-contained: Minimal external dependencies
• RL-ready: Compatible with standard RL frameworks and Gymnasium interfaces

Supported Games

Sokoban

A classic box-pushing puzzle game where the player must push boxes onto target locations.

• Gameplay: Navigate a warehouse, push boxes onto marked targets
• Difficulty: Configurable puzzle complexity (room size, number of boxes)
• Reward Structure: Positive rewards for successful moves, penalties for invalid actions
• State Representation: Text-based grid with symbols for walls, boxes, targets, and player

Tetris

The classic falling-block puzzle game.

• Gameplay: Rotate and place falling tetrominoes to clear lines
• Difficulty: Configurable board size and piece generation
• Reward Structure: Points for line clears, penalties for game over
• State Representation: Text-based board representation

Integration with GRL

These environments are designed to work with the GRL framework.

Integration with NeMo Gym

These environments are in the process of being integrated into NVIDIA NeMo Gym as resource server environments for RL training.

Citation

If you use lmenv in your research, please cite:

@software{lmenv2025,
  title={lmenv: Game Environments for GRL},
  author={lmgame-org},
  year={2025},
  url={https://github.com/lmgame-org/lmenv}
}

Related Projects

• GRL - Game Reinforcement Learning framework
• NVIDIA NeMo Gym - Integration of these environments
• gym-sokoban - Original Sokoban implementation