This repo contains diverse traffic scenarios for evaluating autonomous vehicles in simulation. All scenarios are reactive and can be ran in MetaDrive Simulator.

Currently we provide three sources of traffic scenarios:

• Synthetic Scenarios: Maps are generated via Procedural Generation (PG). Traffic vehicles are generated and controlled according to rules

  Synthetic Dataset Statistics:
    Number of scenarios: 3000
    Number of traffic vehicles per scene: 8.9±3.1
    Block Distrubution:
      Curve: 0.156
      Straight: 0.331
      Roundabout: 0.077
      T-intersection: 0.074
      Intersection: 0.077
      Ramp (merge): 0.081
      Ramp (diverge): 0.076
      Bottleneck (merge): 0.065
      Bottleneck (diverge): 0.068
  

• Real-world Scenarios: These scenarios are built from Waymo motion dataset. Maps and surrounding vehicles are collected and recorded in real world.

  Real-world Dataset Statistics:
    Number of scenarios: 1165
    Number of traffic vehicles per scene: 26.1±21.5
  

• Generated Scenarios: Empowered by our generative model TrafficGen, new traffic flow can be generated and resembles the real-world data given a HD map.

Prerequisite: Install MetaDrive first via:

git clone https://github.com/metadriverse/metadrive.git
cd metadrive
pip install -e .

MetaDrive-Scenario Installation:

git clone [email protected]:metadriverse/metadrive-scenario.git
cd metadrive-scenario
pip install -e .

Download Dataset:

1. Download data from https://github.com/metadriverse/metadrive-scenario/releases.
2. Place data under metadrive-scenario/metadrive_scenario/dataset

We provide an example script:metadrive_scenario/examples/run_scenarios.py, where basic usage and APIs are described. For driving in the synthetic scenarios, run:

python metadrive_scenario/examples/run_scenarios.py  --dataset env_num_3000_start_seed_0_synthetic --scenario_start=0 --scenario_end=3000 

For driving in the real-world Waymo scenarios, run:

python metadrive_scenario/examples/run_scenarios.py  --dataset env_num_1165_waymo --scenario_start=0 --scenario_end=1000 

The scenarios will be built by replaying collected surrounding vehicles' trajectories, while you can add argument --idm_traffic to turn these vehicles into reactive ones.

For both scenario types, you can add the optional argument --manual_control to control the vehicle via w, a, s, d. Also, you can add another argumane --topdown to use 2-D birdeye-view renderer, which is built with pygame.

Refer to Documentation of MetaDrive for detail.

If you use MetaDrive in your own work, please cite:

@article{li2022metadrive,
  title={Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning},
  author={Li, Quanyi and Peng, Zhenghao and Feng, Lan and Zhang, Qihang and Xue, Zhenghai and Zhou, Bolei},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year={2022}
}