The Obstacle Tower environment is a procedurally generated 3D challenge designed for AI agents, testing their capabilities in computer vision, locomotion, planning, and generalization. It offers a platforming-style experience with increasing difficulty, suitable for researchers and developers in reinforcement learning and AI.

How It Works

The environment simulates a tower with procedurally generated floors, rooms, and challenges, requiring agents to adapt to novel combinations of visual appearance, puzzle configurations, and layouts. This design ensures agents must generalize beyond specific training instances to succeed.

Quick Start & Requirements

• Install via pip: pip install -e . after cloning the repository.
• Requirements: Python 3.6+, Unity ML-Agents 1.x, OpenAI Gym.
• The environment binary is automatically downloaded by default. Manual downloads are available for Linux, Mac OS X, and Windows.
• Basic Usage Jupyter Notebook: https://github.com/Unity-Technologies/obstacle-tower-env/blob/master/notebooks/basic_usage.ipynb

Highlighted Details

• Procedurally generated floors with increasing difficulty.
• Tests generalization across vision, control, and planning.
• Supports customization via reset parameters for difficulty and content.
• Includes an evaluation script for benchmarking.

Maintenance & Community

• Project has seen multiple releases and updates, including major version changes for Unity and ML-Agents.
• Benchmarking guidelines and training guides (e.g., for Dopamine on GCP) are provided.

Licensing & Compatibility

• The repository is licensed under the Apache 2.0 license.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The environment is primarily designed for AI agents and requires specific Python and ML-Agents versions. While binaries are provided, users may encounter setup issues depending on their system configuration.