Research map / Reinforcement learning / Visual and representation RL
RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework
Hao Gao, Shaoyu Chen, Yifan Zhu, Yuehao Song, Wenyu Liu, Qian W. L. Zhang, Xinggang Wang
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 |
|---|---|---|---|
| add on-policy generator RL optimization (ablation) | IL-pretrained diffusion planner · BEV-Warp safety-oriented (512 scenarios) | CR: 0.533 -> 0.287 | Improved |
| add on-policy generator RL optimization (ablation) | IL-pretrained diffusion planner · BEV-Warp efficiency-oriented (512 scenarios) | EP@1.0: 0.516 -> 0.391 | Hurt |
| add IL fine-tuning stage during generator optimization (ablation) | generator RL · BEV-Warp efficiency-oriented (512 scenarios) | EP@1.0: 0.391 -> 0.527 | Improved |
| add RL discriminator training alone (ablation) | IL-pretrained planner · BEV-Warp safety-oriented (512 scenarios) | CR: 0.533 -> 0.337 | Improved |
| add discriminator RL to generator RL + IL fine-tuning (full RAD-2) (ablation) | RAD-2 · BEV-Warp safety-oriented (512 scenarios) | CR: 0.403 -> 0.234 | Improved |
| initialize discriminator from planning head instead of random (ablation) | RAD-2 discriminator · BEV-Warp safety-oriented (512 scenarios) | CR: 0.426 -> 0.337 | Improved |
| add reward-variance clip filtering (ablation) | RAD-2 · BEV-Warp efficiency-oriented (512 scenarios) | EP@1.0: 0.662 -> 0.728 | Improved |
| add reward-variance clip filtering (ablation) | RAD-2 · BEV-Warp safety-oriented (512 scenarios) | CR: 0.340 -> 0.337 | No clear effect |
| add entropy regularization to RL objective (ablation) | RAD-2 · BEV-Warp safety-oriented (512 scenarios) | CR: 0.254 -> 0.234 | Improved |
| execution horizon H_reuse 8 vs 2 (ablation) | RAD-2 · BEV-Warp safety-oriented (512 scenarios) | CR: 0.355 -> 0.337 | Mixed |
| TC-GRPO group size 4 vs 2 and 8 (ablation) | RAD-2 · BEV-Warp safety-oriented (512 scenarios) | CR: 0.313 (2), 0.234 (4), 0.291 (8) | Improved |
| increase inference candidates M from 8 to 128 (ablation) | RAD-2 · BEV-Warp efficiency-oriented (512 scenarios) | EP@1.0: 0.667 -> 0.814 | Improved |
Closest papers by meaning
- RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- CarPlanner: Consistent Auto-regressive Trajectory Planning for Large-Scale Reinforcement Learning in Autonomous Driving (CVPR 2025 · similarity 0.88)
- Diffusion-ES: Gradient-free Planning with Diffusion for Autonomous and Instruction-guided Driving (CVPR 2024 · similarity 0.88)
- Diffusion-Based Planning for Autonomous Driving with Flexible Guidance (ICLR 2025 · similarity 0.88)
- ChainFlow-VLA: Causal Flow Planning with Vision-Language Models (NeurIPS 2026 · similarity 0.88)
- Beyond Self-Play and Scale: A Behavior Benchmark for Generalization in Autonomous Driving (NeurIPS 2026 · similarity 0.87)
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.
