Research map / Multi-agent systems and games / Cooperative multi-agent RL
MetaDrive: Composing Diverse Driving Scenarios for Generalizable Reinforcement Learning
Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, Bolei Zhou
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train on more procedurally generated scenes (larger N) | SAC and PPO agents · MetaDrive PG scenes (test) | test success rate | Improved |
| train on more procedurally generated scenes (larger N) | SAC and PPO agents · MetaDrive PG scenes (test) | rule violation and crash rate | Improved |
| increase number of real Waymo training cases | SAC · Waymo real scenarios | test success rate | Improved |
| increase PG training set diversity, tested on real Waymo scenes | PPO · Waymo real test set | test success rate | No clear effect |
| raise the ratio of real scenes in the training mix (0% to 100%) | SAC · Waymo real test set | test success rate | Improved |
| use Lagrangian with PID multiplier instead of reward shaping | SAC · MetaDrive safe exploration | cumulative cost: 3.38 -> 1.90 | Improved |
| use Lagrangian instead of reward shaping | PPO · MetaDrive safe exploration | cumulative cost: 3.33 -> 1.82 | Improved |
| use off-policy SAC instead of on-policy PPO | reward-shaping RL · MetaDrive safe exploration | success rate: 0.207 -> 0.801 | Improved |
| offline RL (CQL) instead of behavior cloning | limited human demonstrations · MetaDrive safe exploration | success rate: 0.01 -> 0.11 | Improved |
| train safe RL on more scenes | PPO-RS and PPO-Lag · MetaDrive safety generalization | test episode cost | Improved |
| concatenate K=4 nearest vehicle states into centralized critic instead of mean-field averaging | CCPPO · MetaDrive MARL Tollgate | success rate (%): 48.62 -> 31.92 | Hurt |
| use CoPO instead of IPPO | multi-agent PPO · MetaDrive MARL Intersection | success rate (%): 73.93 -> 77.79 | Improved |
Closest papers by meaning
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.89)
- IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning (ICLR 2025 · similarity 0.88)
- Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.88)
- Enabling Adaptive Agent Training in Open-Ended Simulators by Targeting Diversity (NeurIPS 2024 · similarity 0.87)
- MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Raw2Drive: Reinforcement Learning with Aligned World Models for End-to-End Autonomous Driving (in CARLA v2) (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
