Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Rethinking Closed-Loop Training for Autonomous Driving
Chris Zhang, Runsheng Guo, Wenyuan Zeng, Yuwen Xiong, Binbin Dai, Rui Hu, Mengye Ren, Raquel Urtasun
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 |
|---|---|---|---|
| use trajectory-based output instead of control output | RAINBOW · Targeted scenarios | pass rate: 0.435 -> 0.814 | Improved |
| use trajectory-based output instead of control output | Imitation learning · Targeted scenarios | pass rate: 0.545 -> 0.617 | Improved |
| use trajectory-based output instead of control output | PPO · Targeted scenarios | pass rate: 0.173 -> 0.273 | Improved |
| use trajectory-based output instead of control output | A3C · Targeted scenarios | pass rate: 0.224 -> 0.362 | Improved |
| add trajectory value learning with model-based counterfactual loss to RAINBOW+T (TRAVL) | RAINBOW+T · Targeted scenarios | pass rate: 0.814 -> 0.865 | Improved |
| add trajectory value learning with model-based counterfactual loss to RAINBOW+T (TRAVL) | RAINBOW+T · Targeted scenarios | collision rate: 0.048 -> 0.026 | Improved |
| train on targeted instead of free-flow scenarios (ablation) | TRAVL · Free-flow test set | pass rate: 0.784 -> 0.903 | Improved |
| train on targeted instead of free-flow scenarios (ablation) | TRAVL · Targeted test set | pass rate: 0.696 -> 0.865 | Improved |
| train on targeted instead of free-flow scenarios (ablation) | RAINBOW+T · Targeted test set | pass rate: 0.453 -> 0.815 | Improved |
| train on targeted instead of free-flow scenarios (ablation) | RAINBOW+T · Free-flow test set | pass rate: 0.783 -> 0.885 | Improved |
| add behavioral variation to map-only variation in training scenarios (ablation) | TRAVL · Targeted scenarios | pass rate: 0.738 -> 0.872 | Improved |
| add map variation to behavioral-only variation in training scenarios (ablation) | TRAVL · Targeted scenarios | pass rate: 0.872 -> 0.865 | No clear effect |
Closest papers by meaning
- RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Safety through feedback in Constrained RL (NeurIPS 2024 · similarity 0.87)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.87)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.86)
- CarPlanner: Consistent Auto-regressive Trajectory Planning for Large-Scale Reinforcement Learning in Autonomous Driving (CVPR 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.
