"格物" (Gewu) is an open-source embodied intelligence simulation platform built on Unity, designed for reinforcement learning research and development. It provides researchers and enthusiasts with an efficient, user-friendly environment to train and deploy AI agents for various robotic tasks, bridging the gap between simulation and real-world application.

How It Works

The platform leverages the Unity game engine for high-fidelity robot simulation. It integrates with Unity ML-Agents for reinforcement learning training pipelines and supports importing diverse robot morphologies via URDF files. For real-world deployment, it offers ROS2 integration enabling Sim2Real transfer, particularly for Ubuntu-based systems. Advanced features include curriculum learning, imitation learning, and VR-assisted control.

Quick Start & Requirements

• Installation: Requires Unity Hub and Unity Editor (version 2022.3). Clone the repository and open the project in Unity. For training, set up a Python 3.10.12 environment (Anaconda recommended) with PyTorch (torch~=2.2.1, cu121 build implies CUDA 12.1) and ML-Agents (mlagents==1.1.0).
• Prerequisites: Cross-platform compatibility (Windows, Linux, macOS). ROS2 Sim2Real requires Ubuntu 20.04/22.04.
• Resources: Unity Editor download is ~7GB. Training setup involves environment configuration.
• Documentation & Demos: Research papers are available. A repository for URDF models is provided. Numerous demo videos showcase platform capabilities, including robot soccer (https://www.bilibili.com/video/BV1NuZBYeEq8/) and Sim2Real (https://b23.tv/y8wXu2N).

Highlighted Details

• Comprehensive Unity-based simulation environment for embodied AI.
• Support for a wide range of robot types (humanoid, quadruped, robotic arms) through URDF import.
• Sim2Real capabilities with ROS2 integration for Ubuntu.
• Advanced functionalities include VR remote operation, large language model (LLM) control, imitation learning, and curriculum learning.
• Includes example scenarios like robot soccer competitions and complex terrain navigation.

Maintenance & Community

Developed collaboratively by major academic institutions including Shanghai University and Tsinghua University. The project shows active development with a roadmap for "Gewu 3.0" featuring VR, LLM integration, and Mujoco engine embedding. Community interaction is facilitated via a WeChat group.

Licensing & Compatibility

The project's license is not explicitly stated in the README, which may impact commercial or closed-source integration. The platform is designed for cross-platform use (Windows, Linux, macOS), with specific ROS2 Sim2Real features requiring Ubuntu.

Limitations & Caveats

The Sim2Real functionality is restricted to Ubuntu operating systems. There is a potential ambiguity regarding hardware requirements for training, with the PyTorch installation command implying CUDA 12.1 while general descriptions suggest CPU-only training is feasible. The absence of a clear license is a significant adoption blocker.