Ngine

Ngine is an open-source framework for building, training, and evaluating embodied AI agents across diverse simulation environments. It provides a unified interface for working with different simulators, robot embodiments, and task domains.

[Note: Ngine is still underconstruction and we are working hard to bring surgical/healthcare benchmarks. stay tuned!]

• Environment Agnostic - Works with multiple simulation backends (Isaac Lab, ManiSkill, PyBullet)
• Plugin Architecture - Extensible design for adding new domains, robots, and simulators
• Multi-Robot Support - Unitree G1/H1, ARX arms, Agilex Piper, Fourier GR1, LeRobot compatible
• Unified MDP Interface - Consistent observation, action, and reward specifications
• VR Teleoperation - Built-in OpenXR support for data collection
• Flexible Asset Loading - Pluggable loaders for local, cloud, or custom backends
• Distributed Training - Scalable infrastructure for parallel environment execution

• OS: Linux (Ubuntu 22.04+ recommended)
• Python: 3.10+
• CUDA: 12.x
• Hardware: NVIDIA RTX GPU
• Isaac Sim: 4.5+ (for Isaac Lab backend)

# Clone repository
git clone https://github.com/Negentropy-ai/Ngine.git
cd Ngine

# Create environment
conda create -n ngine python=3.10
conda activate ngine

# Install package
pip install -e .

# Optional dependencies
pip install -e ".[cloud]"    # Cloud asset loading
pip install -e ".[lerobot]"  # LeRobot integration

# Using shell script
./teleop.sh

# Or directly with config
python ngine/scripts/teleop/teleop_main.py \
    --task_config g1-controller

./train.sh

./eval.sh

# configs/experiment.yml
scene:
  type: robocasa_kitchen
  layout: 1

robot:
  type: g1_dualarm

task:
  type: pick_and_place
  objects:
    - type: bowl
      placement: counter

simulation:
  num_envs: 4
  device: cuda:0

ngine/
├── ngine/
│   ├── assets/              # Asset management system
│   │   ├── loaders/         # Pluggable asset loaders
│   │   └── registry.py      # Central asset registry
│   ├── benchmarks/          # Task definitions (LIBERO, Robocasa)
│   ├── engine/              # Core framework
│   │   ├── cfg/             # Configuration system
│   │   ├── devices/         # Input devices (VR, keyboard)
│   │   ├── embodiments/     # Robot definitions
│   │   ├── mdp/             # MDP components
│   │   ├── models/          # Asset and scene models
│   │   ├── orchestrate/     # Environment orchestration
│   │   ├── scenes/          # Scene definitions
│   │   └── tasks/           # Task base classes
│   ├── rl/                  # RL training configs
│   ├── scripts/             # CLI tools
│   └── utils/               # Helper utilities
└── configs/                 # Configuration files

• Isaac Lab (GPU-accelerated)
• ManiSkill (planned)
• PyBullet (planned)

• Kitchen manipulation (Robocasa, LIBERO)
• Object manipulation
• Custom domains via plugin system

# Run tests
pytest tests/ -v

# Format code
black ngine/
isort ngine/

# Type checking
mypy ngine/

Apache License 2.0 - see LICENSE for details.

Contributions welcome! Please see CONTRIBUTING.md for guidelines.

@software{ngine2025,
  title = {ngine: Environment-Agnostic Embodied AI Infrastructure},
  year = {2025},
  url = {https://github.com/Negentropy-ai/Ngine}
}

Built upon: Isaac Lab, Robocasa, LIBERO