World action models have demonstrated promising performance in robot policy learning, yet their high inference latency remains a critical bottleneck for real-world deployment. Recent approaches have sought to improve inference efficiency by reducing test-time future-frame imagination or accelerating action prediction. However, these methods often leverage representations from all video world model layers without explicitly considering their varying relevance to action prediction, leading to redundant computation and suboptimal representation utilization. In this work, we propose TurboWAM, a fast world action model that efficiently aggregates structured world representations conditioned on action predictions. Specifically, TurboWAM introduces a structured world representation aggregation module that adaptively fuses representations from different world model layers according to their relevance to the evolving action denoising state. Furthermore, TurboWAM employs a structured action supervision mechanism that adaptively modulates gradients from action supervision across different world model layers during backward propagation, facilitating world representation learning. During inference, by reducing the number of world model layers executed and caching relevant representations across action denoising steps, TurboWAM substantially reduces computational overhead. Extensive experiments on multiple robot manipulation benchmarks demonstrate that TurboWAM achieves low inference latency while maintaining competitive task success rates compared with state-of-the-art methods.
Figures & tables
Figure 2: Overview of the TurboWAM framework. The structured world representation aggregation module adaptively fuses representations from different world model layers according to the current action denoising state. The structured action supervision mechanism modulates action-supervision gradients across world model layers during the backward pass. Together, these complementary designs reduce inference latency while maintaining competitive task success rates.
Figure 3: Motivation for the proposed TurboWAM. (a) Distribution of fusion weights across world model layers in Faster-WAM. (b) Spatial feature relationships at different world-model depths in Faster-WAM, visualized through cosine similarity between the reference location (close to the gripper) and other spatial locations. (c) Influence of individual world model layers at different denoising steps in Faster-WAM, measured by the relative change in predicted action velocity when the key/value pairs from a given layer are excluded from fusion. (d) Routing distance across denoising steps, measured by the total variation distance between two layer-wise fusion weight distributions.
Type
Method
EPT
LIBERO
RoboTwin 2.0
Spatial ↑
Object ↑
Goal ↑
Long ↑
Avg ↑
Latency ↓ (ms)
Clean ↑
Rand ↑
Avg ↑
Latency ↓ (ms)
VLA ‡
X-VLA [ 28 ]
✓
98.2
98.6
97.8
97.8
98.1
113.3
72.9
72.8
72.9
125.0
π0.5 [ 3 ]
✓
98.8
98.2
98.0
92.4
96.9
42.0
82.7
76.8
79.8
36.7
WAM
CosmosPolicy [ 11 ]
✓
98.1
100.0
98.2
97.6
98.5
296.9
–
–
–
–
Motus [ 2 ]
✓
96.8
99.8
96.6
97.6
97.7
1555.3
88.7
87.0
87.8
1534.7
LingBot-VA [ 14 ]
✓
98.5
99.6
97.2
98.5
98.5
3032.3
92.9
91.5
92.2
3328.0
Table 1: Comparison with state-of-the-art methods on LIBERO and RoboTwin 2.0. EPT: Embodied Pre-Training. † Implementation based on the arXiv:2608.02365v1 version. ‡ VLA results are for reference only.
Type
Method
EPT
LIBERO-Plus
Camera ↑
Robot ↑
Language ↑
Light ↑
Background ↑
Sensor ↑
Object ↑
Avg ↑
Latency (ms) ↓
VLA ‡
X-VLA [ 28 ]
✓
23.4
89.7
75.7
88.2
96.0
62.7
71.8
71.4
128.9
π0.5 [ 3 ]
✓
75.4
77.5
85.6
96.9
94.6
89.7
85.7
85.7
42.4
WAM
CosmosPolicy [ 11 ]
✓
75.8
63.3
81.7
96.5
88.9
92.7
82.2
82.2
323.8
LingBot-VA [ 14 ]
✓
40.9
83.0
86.4
82.3
53.1
64.4
76.2
69.5
3757.0
Faster-WAM ∗ [ 18 ]
✗
71.6
54.0
89.6
93.6
55.3
81.6
80.2
75.3
84.9
Table 2: Comparison with state-of-the-art methods on LIBERO-Plus across seven perturbation categories. TurboWAM + denotes results obtained with corrected language prompts following https://github.com/sylvestf/LIBERO-plus/issues/64.
Figure 4: Qualitative comparison of task execution. Each row illustrates the progression of task execution from left to right. For clarity, we show only the primary camera view. Additional multi-camera visualizations are provided in the supplementary material.
TurboWAM + [following method]
LIBERO
LIBERO-Plus
Structured aggregation (default)
99.0
78.1
w/o the learned static logit
98.8
77.3
Direct average over layers
98.5
75.9
DoT aggregation [ 18 ]
98.8
77.0
Table 3: Ablation study on structured representation aggregation. Average success rate ↑ is reported.
TurboWAM + [following method]
LIBERO
LIBERO-Plus
Structured supervision (default)
99.0
78.1
Ordinary gradients
97.9
69.2
Gradient isotropic scaling
98.4
75.2
Gradient insulating [ 6 ]
93.9
56.6
Table 4: Ablation study on structured action supervision. Average success rate ↑ is reported.