Direct-policy World Action Models (WAM) have demonstrated the benefits of decoupling action generation from explicit future visual prediction for efficient inference in autonomous driving and embodied control. However, they still require tens of denoising steps to achieve competitive action quality, incurring substantial latency and hindering real-time deployment. Simply reducing the number of denoising steps will collapse the nonlinear multi-step denoising trajectory into coarse updates, substantially degrading action quality. To reduce the denoising steps while enabling real-time action generation, we present an effective world action model with one-step action generation, namely OSWAM, for autonomous driving and embodied control. We argue that it is the absence of structured action priors in random noise initialization that forces direct-policy WAMs to recover coherent action sequences through iterative denoising. Hence, we construct imagination tokens to initialize the diffusion process, providing physical priors and action-conditioned representations. We further introduce privileged information distillation with future conditioning to enhance future-aware action generation. As a result, OSWAM can efficiently and effectively generate high quality actions in just one diffusion step. Our experiments demonstrate that OSWAM achieves comparable or even better results in both autonomous driving and embodied control than direct-policy WAM and Joint-WAM that require tens of steps. It obtains approximately 3× speedup over direct-policy WAM baselines and supports real-time compiled inference with latency as low as 40ms on an NVIDIA RTX 5090 GPU. The project page is available at https://gwenzhang.github.io/OSWAM/
Figures & tables
Figure 1: NAVSIM and Robotwin 2.0 inference efficiency-performance comparison. EPT denotes embodied pretraining. All latencies are measured on NVIDIA H20 GPU.
Figure 2: Quantitative comparison of the Fast-WAM [ 4 ] test-time scaling with different action denoising steps on (a) the autonomous driving dataset Navsim, and (b) embodied control dataset RoboTwin.
Figure 3: The overall architecture of OSWAM. A single Video DiT pass produces an observation-conditioned action prior with AIT and complementary scene context with SIT, providing a structured initialization for one-step Action DiT denoising. Privileged-information distillation injects future-informed representations during training, while inference uses only the current observation.
Figure 4: (a) The attention mask of OSWAM and the privileged teacher. (b) Role-aware RoPE explicitly disentangles token groups using unique positional encodings. sI is SIT and aI is AIT.
Method
Nv
Na
NC ↑
DAC ↑
EP ↑
TTC ↑
C ↑
PDMS ↑
Latency ↓
Human Agent
–
–
100.0
100.0
87.5
100.0
99.9
94.8
–
Traditional E2E and VLA-based planners
UniAD [ 36 ]
–
–
97.8
91.9
78.8
92.9
100.0
83.4
–
TransFuser [ 26 ]
–
–
97.7
92.8
79.2
92.8
100.0
84.0
–
DiffusionDrive [ 8 ]
–
–
98.2
96.2
82.2
94.7
100.0
88.1
–
WoTE [ 31 ]
–
–
98.5
96.8
81.9
94.9
99.9
88.3
–
Table 1: Comparison with state-of-the-art planners on the NAVSIM navtest . Nv and Na denote the number of video and action denoising steps . † SimWAM is reported without reinforcement learning. ‡ DriveVA is reported without CARLA mix training. ∗ Latency reproduced by us.
Method
Nv
Na
EPT
Clean
Rand.
Average
Latency (ms)
VLA-based methods
π0 [ 5 ]
–
–
w/
65.92
58.40
62.20
–
π0.5 [ 6 ]
–
–
w/
82.74
76.76
79.80
–
X-VLA [ 45 ]
–
–
w/
72.90
72.80
72.80
–
Imagine-then-execute-based WAM methods
Motus [ 16 ]
–
–
w/
88.66
87.02
87.80
2367
Table 2: Success rates on RoboTwin 2.0 simulation (Clean and Randomized splits, 50 tasks). EPT denotes embodied pretraining.
Dataset
Model
VAE GPU Time
Visual GPU Time
Action GPU Time
Peak Mem. (GB)
Overall Latency
Compiled Latency
RoboTwin 2.0
LingBot-VA
32.65
1165.12
1205.38
19.12
3365.35
–
Motus
11.14
567.37
Shared
17.95
1472.85
–
Fast-WAM
5.27
18.45
67.61
12.70
323.67
73.64
OSWAM
5.27
23.93
6.78
12.75
95.35
40.81
Navsim
SimWAM
9.83
22.80
72.78
12.87
330.8
90.53
OSWAM
9.82
30.87
7.32
12.90
99.70
52.64
Table 3: Inference efficiency on RoboTwin 2.0 and Navsim. Latency is measured with cached language context on a single NVIDIA RTX 5090 GPU and Intel Xeon Platinum 8470Q. All time are reported in milliseconds (ms).
Table 8Table 9
Fast-WAM
OSWAM (Ours)
Task
SR ↑
Time (s) ↓
SR ↑
Time (s) ↓
Bread into bowl
80
24.6
85
16.2
Eraser into drawer
75
17.6
80
13.8
Table 8: Real-world evaluation on PiPER. SR: success rate (%). Time: video-estimated completion time over successful episodes.
Figure 5: Real-world manipulation with OSWAM on a PiPER robotic arm: (a) placing bread into a bowl and (b) placing a board eraser into an open drawer.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Nv
Na
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 and VLA-based planners
TransFuser [ 26 ]
–
–
96.9
89.9
97.8
99.7
87.1
95.4
92.7
98.3
87.2
76.7
Hydra-MDP++ [ 49 ]
–
–
97.2
97.5
99.4
99.6
83.1
96.5
94.4
98.2
70.9
81.4
GTRS-Dense [ 50 ]
–
–
97.6
97.5
99.0
99.9
87.9
97.0
95.9
97.5
55.9
82.3
DriveSuprim [ 51 ]
–
–
97.5
96.5
99.4
99.6
88.4
96.6
95.5
98.3
77.0
83.1
Appendix
Table A.1: Comparison with state-of-the-art planners on the NAVSIM-v2 navtest . Nv and Na denote the number of video and action denoising steps, respectively.
Method
Auxiliary Supervision
L2 (m) ↓
Collision Rate (%) ↓
1s
2s
3s
Avg.
1s
2s
3s
Avg.
ST-P3 [ 55 ]
Map & Box & Depth
1.33
2.11
2.90
2.11
0.23
0.62
1.27
0.71
UniAD [ 36 ]
Map & Box & Motion & Tracklets & Occ
0.48
0.96
1.65
1.03
0.05
0.17
0.71
0.31
OccNet [ 56 ]
3D-Occ & Map & Box
1.29
2.13
2.99
2.14
0.21
0.59
1.37
0.72
OccWorld [ 30 ]
3D-Occ
0.52
1.27
2.41
1.40
0.12
0.40
2.08
0.87
VAD-Tiny [ 57 ]
Map & Box & Motion
0.60
1.23
2.06
1.30
0.31
0.53
1.33
0.72
Appendix
Table A.2: End-to-end motion planning performance on the nuScenes [ 47 ] dataset.
Table 14
Figure B.1: Qualitative comparison of one-step SimWAM and OSWAM on Navsim v1.
World Action Models (WAMs) unify visual world modeling and action generation, yet their real-world deployment is heavily bottlenecked by iterative action denoising and the sequential execution of visual prefilling and action prediction. We demonstrate that the action-denoising trajectories in WAMs are approximately straight, rendering the first-step velocity a strong estimate of the final action direction. Building on this observation, we introduce 1Step-WAM, a framework that transforms multi-step action denoising into one single forward. Specifically, we employ a lightweight calibrator to directly correct the first step velocity, and use bandwidth-constrained action decoding to further reduce errors in the calibrated one-step actions. Compressing action generation into a single step enables it to execute in parallel with visual prefilling, completely breaking the sequential bottleneck of the traditional imagination-then-action paradigm. Evaluations on RoboTwin and LIBERO show that 1Step-WAM largely preserves multi-step task success with only about 13.4M trainable parameters, while achieving inference speedups of 7.44×, 5.98×, and 24.84× for FastWAM, ImageWAM and Kairos, respectively. Furthermore, real-world robotic deployments using the FastWAM backbone confirm a consistent 7.44× speedup while maintaining close success rate to the multi-step baseline. Code is avaliable at https://github.com/changdong1687/1Step-WAM.
Rundong Su, Pengtao Chen, Haohan Geng +6
Fudan University · University of Science and Technology of China · Peking University +1
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. The code and model weights are available at https://github.com/H-EmbodVis/SimWAM/.
Zongchuang Zhao, Xin Zhou, Tianyang Xu +6
Huazhong University of Science & Technology · Dongfeng Research & Development Institute
A plausible scene evolution depends on the maneuver being considered, while a good maneuver depends on how the scene may evolve. Existing World Action Models (WAMs) largely miss this reciprocity, treating world prediction and action generation as either isolated parallel branches or rigid predict-then-plan pipelines. We formalize this perspective as World-Action Interactive Models (WAIMs), and instantiate it in autonomous driving with DAWN (Denoising Actions and World iNteractive model), a simple yet strong latent generative baseline. DAWN operates in a compact semantic latent space and couples a World Predictor with a World-Conditioned Action Denoiser: the predicted world hypothesis conditions action denoising, while the denoised action hypothesis is fed back to update the world prediction, so that both are recursively refined during inference. Rather than eliminating test-time world evolution altogether or rolling out the full future in pixel space, DAWN performs a short explicit latent rollout that is sufficient to support long-horizon trajectory generation in complex interactive scenes. Experiments show that DAWN achieves strong planning performance and favorable safety-related results across multiple autonomous driving benchmarks. More broadly, our results suggest that interactive world-action generation is a principled path toward truly actionable world models.
Hongbo Lu, Liang Yao, Chenghao He +6
1COWARobot Co. Ltd · 2Shanghai Jiao Tong University · 3Hohai University