README.md

metadata

title: Tetris OpenEnv
emoji: 🎮
colorFrom: purple
colorTo: blue
sdk: docker
app_port: 8000
base_path: /play
tags:
  - openenv

Tetris OpenEnv

A Tetris RL environment for LLM agent training, built on OpenEnv 0.2.1.

LLM agents receive a text-based board representation and must choose spatial actions (left, right, rotate, drop) to play Tetris. Features combo scoring where clearing multiple lines simultaneously gives disproportionately higher rewards.

Problem Statement

Wild Card (#5) - Teaching LLMs spatial reasoning through Tetris. The agent must interpret a 2D text grid and plan piece placements, a fundamentally non-linguistic task solved through language.

Quick Start

from tetris_env import TetrisEnvClient, TetrisAction

with TetrisEnvClient(base_url="https://VortexedSquirrel-tetris-env.hf.space") as env:
    result = env.reset(seed=42)
    while not result.done:
        action = TetrisAction(action="drop")
        result = env.step(action)
        print(f"Reward: {result.reward}, Score: {result.observation.score}")

Actions

Action	Description
left	Move piece left
right	Move piece right
rotate_cw	Rotate clockwise
rotate_ccw	Rotate counter-clockwise
drop	Hard drop to bottom
down	Soft drop one row
noop	Do nothing

Reward Structure

Lines Cleared	Reward	Multiplier
1	+100	x1
2	+300	x3
3	+700	x7
4 (Tetris!)	+1500	x15

Penalties: -1/step, -2height, -5holes, -500 game over.

Built With

• OpenEnv 0.2.1 by Meta PyTorch
• Deployed on Hugging Face Spaces