MetaDrive Documentation

DI-drive is a Decision Intelligence Platform for Autonomous Driving simulation developed under OpenDILab. It is designed to apply Deep Learning methods to decision-making, planning, and control tasks in autonomous driving simulations.

Key features include:

• High ease of use and low barrier to entry.
• Casezoo: A design component intended to make simulation environments closer to real-world driving.
• Simulator Support: It supports both Carla and MetaDrive simulators.
• Deep Learning Integration: It develops driving policies using PyTorch and DI-engine.
• MetaDrive Integration: Provides tutorials for training agents specifically within the MetaDrive simulator and shares pre-trained models.

MetaDrive is an efficient, compositional driving simulator designed for AI and autonomy research, specifically for Reinforcement Learning (RL).

Key features include:

• Compositional Scene Generation: Supports generating infinite scenes with diverse road maps and traffic settings to facilitate research into generalizable RL.
• High Performance: Lightweight design capable of running up to 1,500 FPS on standard PCs (and over 2,000 FPS on MacBook Pro).
• Realistic Simulation: Provides accurate physics and multiple sensory inputs, including:
  • Lidar
  • RGB images
  • Top-down semantic maps
  • First-person view images.

Due to limitations in the underlying simulation engine, each process should only have one single MetaDrive instance.

When parallelizing training environments, you must use process-level parallelization instead of thread-level parallelization.

To run the bridge, you must launch the ROS publishers and the socket server in separate terminal sessions.

1. Terminal 1: Launch the ROS2 publishers using the provided launch file.
2. Terminal 2: Launch the ZMQ socket server using the Python script.

# Terminal 1, launch ROS publishers
ros2 launch metadrive_example_bridge metadrive_example_bridge.launch.py

# Terminal 2, launch socket server
python ros_socket_server.py

MetaDrive uses specific patches for the Panda3D engine to enable advanced render pipeline features that are not yet part of the official Panda3D repository.

Workflow Requirement: After applying a patch to your Panda3D installation, you must delete the patch file from this directory. Deleting the file signals to the MetaDrive render pipeline that the patch has been successfully applied and is available for use.

ScenarioNet is an external ecosystem coupled with MetaDrive. You can access its various components via the following links:

• Webpage: https://metadriverse.github.io/scenarionet/
• Source Code: https://github.com/metadriverse/scenarionet
• Scenario Gallery: https://metadriverse.github.io/snetasset/snet-asset.html
• Official Documentation: https://scenarionet.readthedocs.io/en/latest/

The bridge requires two separate processes running in different terminals to function: one to launch the ROS2 publishers and another to launch the socket server that communicates with MetaDrive.

1. Terminal 1: Launch the ROS2 publishers using the provided launch file.
2. Terminal 2: Launch the socket server script.

# Terminal 1, launch ROS publishers
ros2 launch metadrive_example_bridge metadrive_example_bridge.launch.py

# Terminal 2, launch socket server
python ros_socket_server.py 

To use custom assets such as textures, 3D models, or other external resources within a plugin prefab, place them directly into the metadrive/render_pipeline/rpplugins/plugin_prefab/resources/ directory.

Once placed in this folder, you can retrieve the resource within your code using the get_resource method by providing the filename including its extension.