MetaDrive: Composing Diverse Driving Scenarios for Generalizable Reinforcement Learning.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Science

Background:

• Safe autonomous driving demands generalizability, environmental awareness, and complex decision-making.
• Current Reinforcement Learning (RL) research often isolates these capabilities due to a lack of integrated simulation environments.

Purpose of the Study:

• To introduce MetaDrive, a novel, compositional driving simulation platform.
• To facilitate research in generalizable Reinforcement Learning (RL) algorithms for machine autonomy.

Main Methods:

• MetaDrive generates diverse driving scenarios via procedural generation and real-world data integration.
• Developed single-agent and multi-agent RL tasks and baselines within MetaDrive.
• Benchmarked agent generalizability, safe exploration, and multi-agent traffic learning.

Main Results:

• Increased training data diversity and size significantly improved RL agent generalizability across unseen and real-world scenarios.
• Evaluated and benchmarked various safe RL and multi-agent RL algorithms within MetaDrive.

Conclusions:

• MetaDrive provides a robust platform for advancing autonomous driving research.
• The study demonstrates the critical role of diverse training data in enhancing RL agent generalizability and safety.