Research map / Image generation / Autoregressive and latent generation
KING: Generating Safety-Critical Driving Scenarios for Robust Imitation via Kinematics Gradients
Niklas Hanselmann, Katrin Renz, Kashyap Chitta, Apratim Bhattacharyya, 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 |
|---|---|---|---|
| stop gradients through the indirect path (policy and renderer), use only the direct kinematics path (ablation) | KING · CARLA scenario generation (overall) | Collision rate: 75.83 -> 82.50 | Improved |
| stop gradients through the indirect path, use only the direct path (ablation) | KING · CARLA scenario generation (overall) | Time to 50% collision rate: 16.14 -> 7.78 | Improved |
| stop gradients through the indirect path, use only the direct path (ablation) | KING · CARLA scenario generation (overall) | Seconds per iteration: 3.27 -> 1.90 | Improved |
| replace black-box optimization (CMA-ES) with kinematics-gradient optimization | AIM-BEV attack · CARLA scenario generation (overall) | Collision rate: 68.33 -> 82.50 | Improved |
| increase number of adversarial agents from 1 to 4 | KING · CARLA scenario generation | Collision rate: 86.25 -> 78.75 | Hurt |
| increase number of adversarial agents from 1 to 2 then 4 | CMA-ES · CARLA scenario generation | Collision rate: 67.50 -> 75.00 -> 62.50 | Mixed |
| fine-tune on regular data D_reg only | AIM-BEV · Held-out KING scenarios | Collision rate: 100.00 -> 57.14 | Improved |
| fine-tune on regular data D_reg only | AIM-BEV · Town10 intersections | Collision rate: 17.48 -> 19.51 | Hurt |
| add safety-critical data D_crit to regular data D_reg in fine-tuning (ablation) | AIM-BEV · Held-out KING scenarios | Collision rate: 57.14 -> 28.57 | Improved |
| add safety-critical data D_crit to regular data D_reg in fine-tuning (ablation) | AIM-BEV · Town10 intersections | Collision rate: 19.51 -> 8.13 | Improved |
| fine-tune on D_crit + D_reg instead of no fine-tuning | AIM-BEV · Town10 intersections | Driving Score: 86.74 -> 90.20 | Improved |
| add regular data D_reg to D_crit in fine-tuning (ablation) | AIM-BEV · Town10 intersections | Driving Score: 88.37 -> 90.20 | Improved |
Closest papers by meaning
- LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving (CVPR 2026 · similarity 0.88)
- DriveDPO: Policy Learning via Safety DPO For End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)
- Prioritizing Perception-Guided Self-Supervision: A New Paradigm for Causal Modeling in End-to-End Autonomous Driving (NeurIPS 2025 · similarity 0.87)
- Generating Traffic Scenarios via In-Context Learning to Learn Better Motion Planner (AAAI 2025 · similarity 0.87)
- ReSim: Reliable World Simulation for Autonomous Driving (NeurIPS 2025 · similarity 0.86)
- Trajeglish: Traffic Modeling as Next-Token Prediction (ICLR 2024 · 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.
