SimWAM: A Simple World Action Model for End-to-End Autonomous Driving

Zongchuang Zhao¹, Xin Zhou¹, Tianyang Xu¹, Zhengyang Sun¹, Kaixuan Zhou², Honglin Li², Dingkang Liang^1†, Xiang Bai¹

¹ Huazhong University of Science & Technology
² Dongfeng Research & Development Institute
^† Project leader.

This repository provides the official implementation of SimWAM for the paper A Simple World Action Model for End-to-End Autonomous Driving, including supervised training and action-only reinforcement learning on NAVSIM.

---

📣 News

• 2026.08.19: Released the SimWAM code and weight on the PhysicalAI-AV dataset.
• 2026.08.07: Released the SimWAM paper, code and weight.

---

📄 Abstract

In autonomous driving, World-Action Models (WAMs) have improved end-to-end planning by transferring video dynamics priors to action prediction, but many still couple planning with future-video generation at inference, incurring substantial computational overhead. We present SimWAM, a simple yet effective WAM that leverages future-video prediction solely as a training-time supervision signal. It co-trains a pretrained video expert and a lightweight action expert with joint flow matching. An isolated attention mask keeps action prediction independent of future frames, allowing trajectory prediction without future-frame generation at inference. This design supports multiple pretrained video backbones and independent action-expert scaling within a shared attention interface, while preserving the joint learning objective. Moreover, we apply reinforcement learning to optimize a compositional driving reward beyond trajectory imitation. Experiments show that SimWAM achieves 91.9 PDMS on NAVSIM with a favorable trade-off between accuracy and latency among world-model-based planners, while transferring zero-shot to nuScenes. It also achieves competitive planning accuracy on WOD-E2E and PhysicalAI-Autonomous-Vehicles. These results position SimWAM as a plain yet solid baseline for efficient autonomous driving.

---

🔍 Overview

_{Overview of SimWAM. During training, the video and action DiTs are jointly optimized for future-frame generation and trajectory prediction via shared attention, while the isolated mask prevents the action tokens from accessing future-frame tokens. During inference and reinforcement learning, the model directly predicts trajectories without explicitly predicting future frames.}

• Joint flow-matching co-training. The video expert — a video Diffusion Transformer initialized from Wan2.2-5B together with its video VAE and T5 text encoder — and a lightweight action DiT (hidden size 1024) are co-trained with joint flow matching over future-frame latents and trajectories. Future-video prediction serves as a training-time supervision signal that transfers a traffic-aware dynamics prior into the shared observation representation used for planning.
• Isolated attention mask. Future-frame tokens and action tokens both attend to the current observation latents while remaining mutually invisible, keeping action prediction independent of future frames. This mask is the only structural modification required to isolate the action tokens from future-frame information.
• Direct trajectory prediction at inference. Because the action expert never depends on future-frame tokens, explicit future-frame generation is omitted at deployment: the standalone action DiT directly predicts trajectories without auxiliary motion modules, substantially reducing inference latency.
• Reinforcement learning. The deterministic flow ODE is reformulated as a marginal-preserving SDE, and a group of candidate trajectories per scenario is optimized with FlowGRPO against the compositional NAVSIM PDM reward. RL focuses on the hard navtrain scenarios with the lowest PDMS after imitation learning and updates only the LoRA adapters of the action expert. The same recipe transfers to the Waymo Open Dataset E2E benchmark with the official Rater Feedback Score as the reward.
• Flexibility. The two experts share no weights and interact only through the unified attention interface: the video backbone is replaceable (e.g., LTX-Video, Wan2.1-1.3B, Cosmos2.5, Wan2.2-5B) and the action expert is independently scalable (0.21B–1.02B) without modifying the learning objective or inference pipeline.

---

📈 Performance

_{SimWAM achieves the best PDMS among world-model-based planners, with a favorable trade-off between accuracy and latency on one A100 GPU.}

NAVSIM navtest

Using only a single front camera at 384×672, SimWAM establishes a new state of the art on NAVSIM navtest, surpassing the strongest VLM-based planner SGDrive by 0.8 points and ExploreVLA, which explicitly incorporates future image prediction, by 1.5 points. It further outperforms the imagine-then-act WAMs DriveLaW and DriveVA by 2.8 and 1.0 points, respectively. Moreover, the co-trained video expert achieves the lowest FVD of 36.3, compared with 53.6 for the next-best method.

Method	Reference	Sensors	NC↑	DAC↑	EP↑	TTC↑	C↑	PDMS↑
Human Agent	-	-	100.0	100.0	87.5	100.0	99.9	94.8
Traditional E2E planners
UniAD	CVPR'23	6×C	97.8	91.9	78.8	92.9	100.0	83.4
TransFuser	TPAMI'22	3×C+L	97.7	92.8	79.2	92.8	100.0	84.0
DiffusionDrive	CVPR'25	3×C+L	98.2	96.2	82.2	94.7	100.0	88.1
WoTE	ICCV'25	3×C+L	98.5	96.8	81.9	94.9	99.9	88.3
SeerDrive	NeurIPS'25	3×C+L	98.4	97.0	83.2	94.9	99.9	88.9
VLM-based planners
AutoVLA	NeurIPS'25	3×C	98.4	95.6	81.9	98.0	99.9	89.1
ReCogDrive	ICLR'26	1×C	97.9	97.3	87.3	94.9	100.0	90.8
DriveVLA-W0 (Flow-Matching)	ICLR'26	1×C	98.4	95.3	80.9	95.2	100.0	87.2
DriveVLA-W0 (Query-Base)	ICLR'26	1×C	98.7	99.1	83.3	95.3	99.3	90.2
ExploreVLA	ECCV'26	1×C	98.8	98.4	83.5	96.5	99.9	90.4
SGDrive	CVPR'26	1×C	98.6	97.8	85.8	96.2	100.0	91.1
World-model-based planners
Epona	ICCV'25	1×C	97.9	95.1	80.4	93.8	99.9	86.2
PWM	NeurIPS'25	1×C	98.6	95.9	81.8	95.4	100.0	88.1
DriveLaW	CVPR'26	1×C	99.0	97.1	81.3	96.7	100.0	89.1
DriveVA	ECCV'26	1×C	99.2	97.5	83.5	98.7	100.0	90.9
DriveWAM	arXiv'26	1×C	98.3	98.1	84.3	95.2	100.0	90.1
SimWAM (Ours)	-	1×C	98.5	99.1	86.5	96.1	99.8	91.9

Future-video generation quality

All methods use a 4 s horizon at 2 Hz on navtest.

Metric	SVD	DrivingGPT	Uni-World VLA	PWM	ForgeDrive	DriveDreamer-Policy	SimWAM (Ours)
FVD↓	227.5	142.6	141.8	86.0	69.2	53.6	36.3

Component analysis

Video-expert fine-tuning, future-video supervision, and reinforcement learning contribute complementary gains, improving PDMS by 5.4 points while preserving the simplicity and efficient inference of the standalone action expert.

Configuration	NC	DAC	EP	TTC	PDMS
Action expert only	98.0	95.2	81.0	93.8	86.5
+ Video expert fine-tuning	98.2	96.8	82.6	94.5	88.4
+ Future-video supervision	98.6	98.0	84.0	95.8	90.3
+ RL	98.5	99.1	86.5	96.1	91.9

Qualitative results

_{Qualitative comparison on two navtest scenarios. After reinforcement learning, the ego commits further along the route while staying collision-free within the drivable area.}

NAVSIM v2 (navtest)

We further evaluate SimWAM on NAVSIM-v2, which adopts a reactive simulation protocol and the unified EPDMS metric. In addition to the original PDMS terms, EPDMS considers Driving Direction Compliance (DDC), Traffic Light Compliance (TLC), Lane Keeping (LK), History Comfort (HC), and Extended Comfort (EC). All NAVSIM-v2 experiments are conducted without reinforcement learning, where SimWAM attains the highest EPDMS of 90.2, surpassing DriveLaW by 1.6 points.

Method	Reference	NC↑	DAC↑	DDC↑	TLC↑	EP↑	TTC↑	LK↑	HC↑	EC↑	EPDMS↑
Human Agent	-	100.0	100.0	99.8	100.0	87.4	100.0	100.0	98.1	90.1	90.3
Traditional E2E planners
TransFuser	TPAMI'22	96.9	89.9	97.8	99.7	87.1	95.4	92.7	98.3	87.2	76.7
DiffusionDrive	CVPR'25	98.2	95.9	99.4	99.8	87.5	97.3	96.8	98.3	87.7	84.5
VLM-based planners
ReCogDrive*	ICLR'26	98.3	95.2	99.5	99.8	87.1	97.5	96.6	98.3	86.5	83.6
SGDrive	CVPR'26	98.6	94.3	99.5	99.9	86.0	97.9	96.1	98.3	85.9	86.2
World-model-based planners
DriveVLA-W0	ICLR'26	98.5	99.1	98.0	99.7	86.4	98.1	93.2	97.9	58.9	86.1
DriveLaW	CVPR'26	98.7	96.9	99.6	99.8	87.5	98.3	97.6	98.4	77.4	88.6
SimWAM (Ours)	-	98.6	98.0	99.7	99.9	87.5	98.4	97.9	98.3	84.4	90.2

* indicates training with reinforcement learning.

NAVSIM v2 (navhard)

navhard focuses on safety-critical scenarios and follows a two-stage closed-loop protocol: stage 1 (S1) evaluates the planner on real-world scenarios, while stage 2 (S2) re-evaluates the corresponding synthesized scenarios with reactive traffic agents. Even before reinforcement learning, SimWAM attains the best overall EPDMS of 37.6, surpassing DriveLaW by 7.0 points with leading DAC, DDC, TTC, and LK, especially in the reactive second stage where surrounding agents respond to the ego vehicle.

Method	Reference	Stage	NC↑	DAC↑	DDC↑	TLC↑	EP↑	TTC↑	LK↑	HC↑	EC↑	EPDMS↑
Traditional E2E planners
TransFuser	TPAMI'22	S1	96.2	79.5	99.1	99.5	84.1	95.1	94.2	97.5	79.1	23.1
		S2	77.7	70.2	84.2	98.0	85.1	75.6	45.4	95.7	75.9
DiffusionDrive	CVPR'25	S1	96.8	86.0	98.8	99.3	84.0	95.8	96.7	97.6	79.6	27.5
		S2	80.1	72.8	84.4	98.4	85.9	76.6	46.4	96.3	72.8
VLM-based planners
SGDrive	CVPR'26	S1	95.8	87.6	97.8	99.8	84.4	94.7	92.9	97.8	28.9	25.5
		S2	79.4	65.4	79.1	98.9	88.9	75.3	42.7	96.4	29.6
ReCogDrive	ICLR'26	S1	96.4	78.9	98.7	99.8	82.6	95.6	94.4	97.6	74.2	25.7
		S2	80.2	65.0	82.4	98.7	85.2	76.9	43.8	96.6	71.8
World-model-based planners
DriveVLA-W0	ICLR'26	S1	96.8	83.3	99.0	99.6	84.6	95.3	96.4	97.6	78.2	24.4
		S2	76.8	64.3	79.9	98.3	89.2	75.0	46.8	95.8	53.1
DriveLaW	CVPR'26	S1	97.3	89.1	99.2	99.6	84.3	97.1	96.2	97.8	67.6	30.6
		S2	82.5	67.6	83.5	98.1	84.8	78.5	45.8	96.4	57.3
SimWAM (Ours)	-	S1	98.0	92.0	99.7	99.6	83.8	96.2	97.3	97.8	71.6	37.6
		S2	81.8	78.6	87.3	98.4	86.3	78.6	49.5	96.3	69.9

Waymo Open Dataset E2E (WOD-E2E)

Open-loop planning on the rater-annotated validation and test splits. ADE denotes the average displacement error and RFS the official Rater Feedback Score. Reinforcement learning is performed on the validation split only; * indicates the RL-fine-tuned model.

Validation split:

Method	ADE@3s↓	ADE@5s↓	RFS↑
Human Driver	-	-	8.13
VAD	3.19	5.81	4.45
UniAD	6.50	10.81	5.78
RAP-DINO	0.97	2.20	7.91
SimWAM (Ours)	0.98	2.29	7.99
SimWAM (Ours)*	0.98	2.27	8.29

Test split:

Method	ADE@3s↓	ADE@5s↓	RFS↑
UniPlan	1.31	2.99	7.78
dVLM-AD	1.29	3.02	7.63
HMVLM	1.33	3.07	7.74
AutoVLA*	1.35	2.96	7.56
NoRD*	1.25	-	7.71
SimWAM (Ours)	1.22	2.69	7.77
SimWAM (Ours)*	1.21	2.68	7.84

Although RL is performed on the validation split, the imitation-trained model alone already surpasses the strongest baseline RAP-DINO there with an RFS of 7.99. On the held-out test split, the imitation-trained model already surpasses all compared methods in ADE, and RL generalizes to this split as well, improving RFS from 7.77 to 7.84.

PhysicalAI-Autonomous-Vehicles

We evaluate on the same 1,000-clip test subset adopted by DriveWAM for a consistent comparison, reporting Average Displacement Error (ADE) and Final Displacement Error (FDE) over 3-second and 4-second future trajectories. Although trained on only 65K samples, SimWAM achieves the best ADE and FDE at both horizons with the imitation-trained model alone. SV denotes single-view camera.

Method	Source	Sensors	Params.	ADE@3s↓	FDE@3s↓	ADE@4s↓	FDE@4s↓
VaVAM*	Valeo	SV	1.3B	2.31	4.32	-	-
Alpamayo-1.5	NVIDIA	SV	10B	0.80	2.31	1.44	4.18
DriveWAM	-	SV	5B + 8B	0.47	1.35	0.83	2.47
SimWAM (Ours)	-	SV	6B	0.40	1.08	0.69	1.96

* evaluated using the released checkpoint, which only supports up to 3s prediction.

Zero-shot generalization on nuScenes

The NAVSIM-trained model is evaluated on the nuScenes open-loop planning benchmark without fine-tuning. * denotes using only the front camera as input.

Method	Finetune	Input	Auxiliary Supervision	L2 1s↓	L2 2s↓	L2 3s↓	L2 Avg.↓	Coll. 1s↓	Coll. 2s↓	Coll. 3s↓	Coll. Avg.↓
ST-P3	✓	Camera	Map&Box&Depth	1.33	2.11	2.90	2.11	0.23	0.62	1.27	0.71
UniAD	✓	Camera	Map&Box&Motion	0.48	0.96	1.65	1.03	0.05	0.17	0.71	0.31
VAD-Base	✓	Camera	Map&Box&Motion	0.54	1.15	1.98	1.22	0.04	0.39	1.17	0.53
GenAD	✓	Camera	Map&Box&Motion	0.36	0.83	1.55	0.91	0.06	0.23	1.00	0.43
Epona	✓	Camera*	None	0.61	1.17	1.98	1.25	0.01	0.22	0.85	0.36
DriveVA	✗	Camera*	None	0.33	0.76	1.43	0.84	0.00	0.07	0.12	0.06
DriveWAM	✗	Camera*	None	0.28	0.81	1.80	0.96	0.00	0.05	0.14	0.06
SimWAM (Ours)	✗	Camera*	None	0.29	0.82	1.77	0.96	0.00	0.03	0.11	0.05

Despite using neither nuScenes supervision nor auxiliary annotations, SimWAM achieves an average L2 error and collision rate comparable to or better than those of strong zero-shot baselines.

---

⚙️ Installation

Clone the repository:

git clone https://github.com/H-EmbodVis/SimWAM.git
cd SimWAM

Create a Python 3.10 environment and install the pinned runtime dependencies:

conda create -n simwam python=3.10 -y
conda activate simwam

python -m pip install -r requirements.txt
python -m pip install -e navsim --no-deps
python -m pip install -e navsim_v2 --no-deps
python -m pip install -e . --no-deps

For NAVSIM, nuPlan, maps, and dataset preparation, follow the official NAVSIM v1.1 repository and the nuPlan devkit. The required release subset is included under navsim/.

All commands are intended to run from the repository root and use relative paths by default.

---

📦 Preparation

ActionDiT initialization

bash scripts/model_prepare.sh

Text embeddings

bash scripts/precomput_text_embed.sh

Use +overwrite=false to keep existing embeddings.

Model weights

Download the checkpoints from the Hugging Face repository into weights/. Every RL checkpoint is LoRA-merged, so the evaluation scripts load it without the LoRA code path.

File	Dataset	Stage
SimWAM.pt	NAVSIM	supervised (main released checkpoint)
SimWAM-IL-step044400.pt	NAVSIM	supervised, step 44400
SimWAM-RL.pt	NAVSIM	FlowGRPO on the hard navtrain subset, step 20000
SimWAM-PAI-AV.pt	PhysicalAI-AV	supervised
SimWAM-Waymo-IL.pt	Waymo E2E	supervised, step 52140
SimWAM-Waymo-RL.pt	Waymo E2E	RFS GRPO, step 2400

---

🏋️ Training and Evaluation

NAVSIM

Supervised training

NNODES=4 \
NPROC_PER_NODE=8 \
bash scripts/train_navsim_zero1_torchrun.sh \
  task=navsim_uncond_front_384x672_1e-4 \
  num_workers=8

FlowGRPO LoRA fine-tuning

RL reformulates the deterministic flow ODE as a marginal-preserving SDE, samples a group of candidate trajectories per scenario, and optimizes only the action expert's LoRA adapters against the compositional NAVSIM PDM reward. It focuses on the hard navtrain subset whose base-model PDM score is below 0.9, and is warm-started from the supervised checkpoint.

NNODES=1 NPROC_PER_NODE=8 bash scripts/train_navsim_grpo_zero1_torchrun.sh \
    task=navsim_grpo_action_pdm_384x672_flowgrpo_lora \
    num_workers=8 \
    output_dir='./runs_grpo/navsim_pdm_below0p9/' \
    data.train.scene_filter="./navsim/navsim/planning/script/config/common/train_test_split/scene_filter/navtrain_pdm_score_below0p9.yaml" \
    grpo.eval.num_batches=8 \
    max_steps=170000 save_every=1000

The task preset already selects that scene filter and 8 eval batches, so the minimal form is equivalent:

NPROC_PER_NODE=8 bash scripts/train_navsim_grpo_zero1_torchrun.sh \
  task=navsim_grpo_action_pdm_384x672_flowgrpo_lora

configs/task/navsim_grpo_action_pdm_384x672_flowgrpo_lora.yaml holds the whole recipe: 384×672 front camera, 8 candidates per condition, 10 denoising steps of which 3 are stochastic, PPO clip 0.005, 4 inner updates per rollout, LoRA r=16 on the action expert's q/k/v/o, and per-dimension KL anchor weights [x, y, heading] = [0.2, 1.0, 1.0]. Shared defaults live in configs/train_grpo.yaml.

The supervised warm start defaults to ./weights/SimWAM.pt; point SIMWAM_IL_CHECKPOINT or model.checkpoint_path=... at your own run to change it. The launcher requires the PDM metric cache at data/metric_cache_navtrain (override with NAVSIM_METRIC_CACHE_PATH).

Saved weights under checkpoints/weights/ are LoRA-merged, so every evaluation script consumes them without the LoRA code path. checkpoints/state/ keeps the adapters, the frozen IL reference and the optimizer for resume=.

Evaluation

CKPT=./runs_grpo/navsim_pdm_below0p9/checkpoints/weights/step_XXXXXX.pt \
TASK=navsim_grpo_action_pdm_384x672_flowgrpo_lora \
NPROC_PER_NODE=8 \
bash experiments/navsim/run_eval_navsim.sh

To verify the released supervised checkpoint with a one-sample smoke test:

CKPT=./weights/SimWAM.pt \
TASK=navsim_uncond_front_384x672_1e-4 \
NPROC_PER_NODE=1 \
bash experiments/navsim/run_eval_navsim.sh \
  EVALUATION.max_samples=1 \
  EVALUATION.num_inference_steps=2 \
  EVALUATION.save_videos=false

To evaluate the released reinforcement-learning checkpoint:

CKPT=./weights/SimWAM-RL.pt \
TASK=navsim_grpo_action_pdm_384x672_flowgrpo_lora \
NPROC_PER_NODE=8 \
bash experiments/navsim/run_eval_navsim.sh

Training outputs are written to runs/; evaluation outputs are written to evaluate_results/navsim/.

The supplied launchers use DeepSpeed ZeRO-1 through scripts/ds_configs/ds_zero1_config.json. The matching Accelerate configuration is scripts/accelerate_configs/accelerate_zero1_ds.yaml.

NAVSIM v2 prediction and scoring

The v2 prediction scripts write token-level trajectories for navtest or navhard_two_stage; the scoring launcher consumes those .npy files with the v2 PDM evaluator. Set CKPT, EXP_NAME, and the dataset/cache roots as needed.

CKPT=./weights/SimWAM.pt EXP_NAME=simwam_v2 \
  bash experiments/navsim/run_predict_navsim_v2.sh

CKPT=./weights/SimWAM.pt EXP_NAME=simwam_navhard \
  bash experiments/navsim/run_predict_navhard.sh

EXP_NAME=simwam_v2 SPLIT=both \
  bash navsim_v2/scripts/evaluation/run_npy_trajectory_agent_pdm_score_evaluation.sh

PhysicalAI

The PhysicalAI data is built from the NVIDIA PhysicalAI-Autonomous-Vehicles dataset. The training/test clip splits and data processing follow DriveWAM (Hugging Face).

The migrated metadata files are data/physicalai_train.jsonl and data/physicalai_dataset_stats.json. Keep the referenced front-camera frames under data/physicalai/images/ using the relative paths in the JSONL file.

The PhysicalAI checkpoint and training data are available on our Hugging Face repository.

Train and evaluate with:

NPROC_PER_NODE=2 bash scripts/train_physicalai_zero1_torchrun.sh \
  task=physicalai_uncond_front_384x672_1e-4

CKPT=./weights/SimWAM-PAI-AV.pt NPROC_PER_NODE=1 \
  bash experiments/physicalai/run_eval.sh \
  EVALUATION.max_clips=1 EVALUATION.num_inference_steps=2

Waymo Open Dataset E2E

SimWAM also trains and fine-tunes on the Waymo Open Dataset End-to-End Driving benchmark: one FRONT camera at 480×512, native 20×2 XY actions (5 s at 4 Hz) and [vx, vy, ax, ay, command(4)] proprioception. See experiments/waymo/README.md for the full data layout and every entry point.

The metadata files are data/waymo_training_front_video4s_traj5s_4hz_xy.jsonl (333,537 IL rows), data/waymo_val_front_current_traj5s_4hz_xy_pref_only.jsonl (479 scored-reference rows, used for validation and for RL) and data/waymo_test_front_current_input.jsonl (1,505 label-free test rows), plus data/waymo_dataset_stats.json for action normalization. Keep the camera frames under data/waymo/ — every image path in the manifests is relative to that root (override with WAYMO_DATA_ROOT), e.g. images/training/<scene_id>/FRONT/009.jpg. The manifests are distributed through our Hugging Face repository; data/waymo_dataset_stats.json and the .summary.json provenance sidecars are tracked in the repository.

Supervised training and evaluation

NPROC_PER_NODE=8 bash scripts/train_waymo_zero1_torchrun.sh \
  task=waymo_uncond_front_512x480_1e-4

CKPT=./weights/SimWAM-Waymo-IL.pt NPROC_PER_NODE=8 \
  bash experiments/waymo/run_eval_waymo_action_only.sh

Validation runs action-only inference over all 479 scored rows and reports ADE/FDE at 1/3/5 s and the official Rater Feedback Score.

RFS reinforcement learning

The same RL recipe used on NAVSIM, with the vendored official RFS as the reward and per-dimension KL weights [x, y] = [0.2, 1.0]:

NPROC_PER_NODE=8 bash scripts/train_waymo_grpo_zero1_torchrun.sh \
  task=waymo_grpo_rfs num_epochs=10 max_steps=null

The recipe, the reward flow and the epoch accounting are documented in experiments/waymo/GRPO.md.

Test-set prediction

CKPT=./weights/SimWAM-Waymo-RL.pt NPROC_PER_NODE=8 \
  bash experiments/waymo/run_predict_waymo_test.sh

make_test_submission.py then packages predictions.jsonl into the official binproto shards and tar.gz; see experiments/waymo/TEST_SUBMISSION.md.

---

👍 Acknowledgement

SimWAM builds upon the following projects and resources:

• NAVSIM for the planning benchmark and evaluation tooling.
• nuPlan / OpenScene for the driving datasets.
• Wan2.2 for the pretrained video generation backbone.
• NVIDIA PhysicalAI-Autonomous-Vehicles for the PhysicalAI driving dataset.
• DriveWAM for the PhysicalAI clip splits and data processing.
• Waymo Open Dataset for the End-to-End Driving benchmark and the official Rater Feedback Score implementation.

---

📖 Citation

If SimWAM is useful in your research, please consider citing the paper:

@article{zhao2026simwam,
  title={SimWAM: A Simple World Action Model for End-to-End Autonomous Driving}, 
  author={Zongchuang Zhao and Xin Zhou and Tianyang Xu and Zhengyang Sun and Kaixuan Zhou and Honglin Li and Dingkang Liang and Xiang Bai},
  journal={arXiv preprint arXiv:2608.07468},
  year = {2026}
}

---

License

See LICENSE. NAVSIM, Wan2.2, nuPlan, OpenScene, and Waymo Open Dataset retain their own licenses and distribution terms. Refer to the official NAVSIM license, Wan2.2 repository, nuPlan license, and Waymo Open Dataset license for upstream terms.

The vendored Rater Feedback Score implementation under src/simwam/datasets/waymo/_vendor/ is an unmodified copy of the Waymo Open Dataset code and stays under its Apache-2.0 license (see the accompanying LICENSE file in that directory).