NegWAM

Reinforcing End-to-End Autonomous Driving Safety via Negative-Aware Planning in a Unified World-Action Model

Paper · Code & Setup · Model Files

NegWAM is a safety post-training framework for unified world-action models in end-to-end autonomous driving. It learns from near-expert failure trajectories: unsafe alternatives that remain geometrically close to expert trajectories. Repulsive anchors steer predictions away from these alternatives, while scene-conditioned rider tokens learn scene–trajectory compatibility. The auxiliary rider branch is used during training and omitted from the inference path.

This repository provides the NegWAM Transformer weights, model configuration, and NAVSIM v2 evaluation results. Use the custom model implementation and evaluation tools in the GitHub repository to run the model.

Model Details

Property	Value
Task	End-to-end autonomous driving / ego-trajectory planning
Model class	WanTransformer3DNavSimModel (custom Diffusers-based implementation)
Conditioning	Camera observations, ego state/history, and navigation commands
Planning output	8 ego poses (x, y, heading) over 4 seconds at 0.5-second intervals through the NAVSIM agent
Weight format	Safetensors, BF16
Weight file size	Approximately 12.49 GB
License	Apache 2.0

Repository Contents

File	Description
config.json	Architecture configuration for the custom Transformer.
diffusion_pytorch_model.safetensors	NegWAM Transformer weights.
navhard_result.csv	NAVSIM v2 navhard two-stage evaluation results and summary rows.
navtest_result.csv	NAVSIM v2 navtest evaluation results and the overall summary row.

Keep config.json and diffusion_pytorch_model.safetensors in the same directory. The evaluation configuration's checkpoint_path must point to that directory.

Included Results

Benchmark	Split / protocol	EPDMS ↑	Result file
NAVSIM v2	navhard, two-stage	40.4	navhard_result.csv
NAVSIM v2	navtest	93.3	navtest_result.csv

Scores are shown on a 0–100 scale and rounded to one decimal place. They come from the extended_pdm_score_combined summary row for navhard and the average_all_frames summary row for navtest. For navhard, use the combined summary score rather than averaging all rows in the CSV, which contains both evaluation stages and aggregate rows.

The paper also reports zero-shot closed-loop evaluation on HUGSIM. See the paper for the protocol and results, and the GitHub release roadmap for code availability.

License and Citation

This release uses the Apache License 2.0.

If you use NegWAM in your research, please cite:

@misc{zhang2026negwam,
  title = {Reinforcing End-to-End Autonomous Driving Safety via Negative-Aware Planning in a Unified World-Action Model},
  author = {Zihao Zhang and Xinyu Zhang and Bing Zeng and Xingyu Chen and Zhenning Li and Bin Yu and Ziyuan Pu},
  year = {2026},
  eprint = {2611.08111},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.08111}
}