Driving · 2026-11-06
Reinforcing End-to-End Autonomous Driving Safety via Negative-Aware Planning in a Unified World-Action Model
Zihao Zhang, Xinyu Zhang, Bing Zeng, Xingyu Chen, Zhenning Li, Bin Yu, Ziyuan Pu
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
NegWAM is a safety post-training method for unified world–action models in end-to-end autonomous driving. It mines near-expert failure trajectories offline, constructs geometrically close but unsafe negatives, and uses them as repulsive anchors plus scene-conditioned rider tokens to learn scene–trajectory compatibility. The auxiliary rider branch is removed at inference. On NAVSIM v2, it reports 40.4 EPDMS on navhard and 93.3 EPDMS on navtest.
Why it matters
The authors target a limitation of world–action models trained mainly by flow matching on expert trajectories: near-expert unsafe alternatives are not explicitly rejected, so explicit safety boundaries may be missing. NegWAM uses offline-mined failure trajectories to introduce negative-aware objectives during post-training, while maintaining an inference path with no added modules (auxiliary branch removed).
Method
- Post-train a unified world–action model with flow matching plus near-expert negative objectives built from offline-mined safety failures.
- Construct negative trajectories that are geometrically close to the expert but have safety risks; treat them as repulsive anchors during action generation.
- Encode expert and negative trajectories as scene-conditioned rider tokens for compatibility learning; remove the auxiliary rider branch at inference.
Limitation
The learned safety boundary depends on the fixed trajectory vocabulary used for near-expert failures; rare maneuvers/interactions outside the candidate space receive no direct negative supervision. The offline failure labels also depend on the evaluator and on ADE measuring geometric proximity rather than interaction risk; the paper notes no formal safety guarantee and requires independent closed-loop validation.
Abstract (from arXiv)
World-action models (WAMs) jointly model future traffic states and driving actions for end-to-end autonomous driving in intelligent transportation systems. However, their action learning remains dominated by expert-trajectory imitation, leaving unsafe alternatives close to expert trajectories insufficiently constrained. To improve WAM robustness in safety-critical scenarios, we propose NegWAM, a safety post-training framework for unified WAMs that learns explicit safety boundaries from near-expert failure trajectories. NegWAM constructs negative trajectories that are geometrically close to the expert trajectory yet carry safety risks, uses them as repulsive anchors to steer predictions away from dangerous behavior, and encodes them as scene-conditioned rider tokens to learn scene--trajectory compatibility. During training, NegWAM jointly optimizes the flow-matching and near-expert negative objectives, while the auxiliary rider branch is discarded at inference, incurring no additional overhead. On NAVSIM v2, NegWAM achieves 40.4 EPDMS on navhard, a 4.1% relative improvement over the strongest compared scorer-free WAM baseline, and achieves 93.3 EPDMS on navtest. These results demonstrate that NegWAM achieves strong performance in safety-critical scenarios while preserving routine-driving performance. Zero-shot closed-loop evaluation on HUGSIM further demonstrates cross-benchmark generalization. The code is available at https://github.com/Zizizi-hao/NegWAM.
Related papers
- Breaking survivorship bias in driving world models
- SafeDriveVLA: Navigation-Conditioned World Model Dreaming for Conflict-Aware End-to-End Autonomous Driving
- Closed-Loop Aware Planning via Imitation-Free Speed Scoring for End-to-End Autonomous Driving
- GuidedWAM: Training-Free Energy Guidance for World Action Models in Autonomous Driving
- BridgeGuard: Explicit Safety Drift for Diffusion-based Autonomous Driving
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-11-05 · Indexed by Pub-AI: 2026-11-06 UTC
Reinforcing End-to-End Autonomous Driving Safety via Negative-Aware Planning in a Unified World-Action Model
Preprint posted: 2026-11-05 · Original record ↗
- arxiv
- 2611.08111
Explicit source versions (1)
- v1 · 2026-11-05
- Preprint available
Source date: 2026-11-05 · First detected: 2026-11-06 UTC · Historical baseline
Source evidence (3)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.08111v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
World-action models (WAMs) jointly model future traffic states and driving actions for end-to-end autonomous driving in intelligent transportation systems. However, their action learning remains dominated by expert-trajectory imitation, leaving unsafe alternatives close to expert trajectories insufficiently constrained. To improve WAM robustness in safety-critical scenarios, we propose NegWAM, a safety post-training framework for unified WAMs that learns explicit safety boundaries from near-expert failure trajectories. NegWAM constructs negative trajectories that are geometrically close to the expert trajectory yet carry safety risks, uses them as repulsive anchors to steer predictions away from dangerous behavior, and encodes them as scene-conditioned rider tokens to learn scene--trajectory compatibility. During training, NegWAM jointly optimizes the flow-matching and near-expert negative objectives, while the auxiliary rider branch is discarded at inference, incurring no additional overhead. On NAVSIM v2, NegWAM achieves 40.4 EPDMS on navhard, a 4.1% relative improvement over the strongest compared scorer-free WAM baseline, and achieves 93.3 EPDMS on navtest. These results demonstrate that NegWAM achieves strong performance in safety-critical scenarios while preserving routine-driving performance. Zero-shot closed-loop evaluation on HUGSIM further demonstrates cross-benchmark generalization. The code is available at https://github.com/Zizizi-hao/NegWAM.
