SPEAR is a simulator designed to address limitations in embodied AI research, specifically offering enhanced content diversity, physical interactivity, and visual fidelity compared to existing tools. It targets researchers and developers in embodied AI, providing a photorealistic and highly detailed environment for training agents.

How It Works

SPEAR leverages Unreal Engine to create its virtual environments, featuring 300 unique indoor scenes with professionally designed layouts and object placements. This approach ensures high visual fidelity and content diversity, avoiding the repetition often found in other simulators. The simulator provides an OpenAI Gym interface, enabling seamless interaction with the environments through Python.

Quick Start & Requirements

• Precompiled binaries are available for download from the release notes.
• System requirements are aligned with Unreal Engine specifications.
• Refer to the Getting Started tutorial for interaction and Python interface details.

Highlighted Details

• Features 300 unique indoor environments with 2,566 rooms and 17,234 individually manipulable objects.
• Environments are designed by professional artists for maximum diversity and realism.
• Provides an OpenAI Gym interface for Python-based agent interaction.

Maintenance & Community

SPEAR was developed with support from Intel and Kujiale, and is now maintained by an independent consortium with continued support from Kujiale.

Licensing & Compatibility

The code is released under an MIT License. Assets are licensed under various licenses, including CC0 for apartment, debug, and warehouse scenes, permitting academic use.

Limitations & Caveats

While the code is MIT licensed, the asset licenses vary, which may require careful review for specific commercial applications or closed-source integration. The README does not detail specific hardware recommendations beyond general Unreal Engine requirements.