Research map / Multi-agent systems and games / Human-AI and goal-conditioned
SLEDGE: Synthesizing Driving Environments with Generative Models and Rule-Based Traffic
Kashyap Chitta, Daniel Dauner, Andreas Geiger
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 rasterized latent map (RLM) instead of rasterized state image (RSI) for lane graph reconstruction (ablation) | lane graph representation · nuPlan lane graph reconstruction, GEO | F1: 0.933 -> 0.980 | Improved |
| add channel group masking (split lane and agent latent groups) (ablation) | RVAE · nuPlan lane graph reconstruction, GEO | F1: 0.981 -> 0.980 | No clear effect |
| use RLM instead of RSI as the diffusion representation (ablation) | DiT-L · nuPlan lane graph generation | Frechet Reach: 1.90 -> 0.88 | Improved |
| scale DiT-L to DiT-XL with about 6x more compute (ablation) | SLEDGE DiT · nuPlan lane graph generation | Frechet Reach: 0.88 -> 0.20 | Improved |
| use latent diffusion DiT instead of HDMapGen autoregressive or VAE baselines | lane graph generation · nuPlan lane graph generation | Frechet Connectivity: 7.02 -> 2.35 (DiT-L RLM vs HDMapGen) | Improved |
| increase training compute budget for DiT with RLM (ablation) | DiT · nuPlan lane graph generation | generation quality | Improved |
| increase training dataset size from 0.25x to 1x (ablation) | DiT with RLM · nuPlan lane graph generation | generation quality | No clear effect |
| extend route length from 100m to 500m in replay | PDM-Closed · SLEDGE Lane->Agent | planner failure rate: 0.06 -> 0.26 | Hurt |
| use hard routes and hard traffic in generated 500m scenes | PDM-Closed · SLEDGE Lane & Agent | planner failure rate: 0.39 -> 0.49 | Hurt |
| use generated easy routes and traffic instead of log replay | PDM-Closed · SLEDGE Lane->Agent 100m | planner failure rate: 0.06 -> 0.07 | No clear effect |
Closest papers by meaning
- Scenario Dreamer: Vectorized Latent Diffusion for Generating Driving Simulation Environments (CVPR 2025 · similarity 0.90)
- SceneDiffuser++: City-Scale Traffic Simulation via a Generative World Model (CVPR 2025 · similarity 0.88)
- BehaviorGPT: Smart Agent Simulation for Autonomous Driving with Next-Patch Prediction (NeurIPS 2024 · similarity 0.87)
- SMART: Scalable Multi-agent Real-time Motion Generation via Next-token Prediction (NeurIPS 2024 · similarity 0.86)
- LASIL: Learner-Aware Supervised Imitation Learning For Long-term Microscopic Traffic Simulation (CVPR 2024 · similarity 0.86)
- Advancing Multi-agent Traffic Simulation via R1-Style Reinforcement Fine-Tuning (ICLR 2026 · 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.
