Xenoverse

English | Chinese

Xenoverse is a collection of procedurally generated environments for benchmarking general decision-making, in-context adaptation, meta-training, and open-world evaluation. Instead of relying on one fixed benchmark family, the repository groups multiple randomized world generators under a single Python package, xenoverse.

Overview

Why Xenoverse

• Diversity over memorization: agents are evaluated across many world-generation processes rather than one static task set.
• Open-ended evaluation: procedural tasks reduce benchmark overfitting.
• Reusable infrastructure: environment families share one namespace and similar task-sampling workflows.
• Agent-oriented extensions: the repository also includes chemistry-agent and town-style environment workspaces.

Main Modules

Repository Layout

Xenoverse/
  README.md
  README.zh.md
  setup.py
  requirements.txt
  xenoverse/

• xenoverse/ is the installable Python package.
• Package code and package-level documentation live under xenoverse/.
• The current package includes anymdp, linds, anyhvac, metalang, mazeworld, metacontrol, chemverse, ai_town_env, and utils.

Installation

Install from PyPI:

pip install xenoverse

Install from source:

git clone https://github.com/FutureAGI/xenoverse
cd xenoverse
pip install .

Core dependencies include:

• gymnasium>=1.0.0
• numpy>=1.24.4
• Pillow>=6.2.2
• six>=1.12.0
• pygame>=2.6.0
• numba>=0.58.1
• scipy

Quick Start

import gymnasium as gym
import xenoverse.anymdp

env = gym.make("anymdp-v0", max_steps=5000)
observation, info = env.reset()

Chemverse

xenoverse.chemverse is the chemistry-agent environment family in this repository. It provides procedurally generated chemistry worlds, tool-based interaction, and backend-driven task evaluation.

Quick entrypoints:

python -m xenoverse.chemverse.demo
python -m xenoverse.chemverse.generate_worlds --list
python -m xenoverse.chemverse.tests.test_backend

See xenoverse/chemverse/README.md for module structure, backend usage, scoring, and world generation details.

Research Positioning

Xenoverse is well suited for:

• meta-reinforcement learning
• in-context reinforcement learning
• open-world evaluation
• domain-randomized control
• procedural task-distribution generalization
• agentic scientific workflows over generated environments

Package Documentation

For the current package structure, environment families, module entrypoints, and package documentation index, see xenoverse/README.md.