Faster-WAM: Do World Action Models Need Deep Action Modules?
2026-08-03 • Artificial Intelligence
Artificial IntelligenceMachine LearningRobotics
AI summaryⓘ
The authors identified that current robot action prediction systems are slow because their action modules are tightly linked with large video processing models. They proposed a new design called Dock of Transformer (DoT), which treats a video model as a hub and connects smaller action modules to it flexibly. They created Faster-WAM using this idea, which is much faster and still performs well on robot action tasks without extra training. Their approach also generalizes better to new situations and reduces the time needed for predictions.
World Action Models (WAMs)Video TransformersAction PredictionInference LatencyDocking InterfaceRoPE RealignmentOut-of-Distribution GeneralizationLIBERO DatasetRoboTwin 2.0
Authors
Liheng Ma, Rui Heng Yang, Zhanguang Zhang, Mateo Clemente, Ziwen Hu, Tongtong Cao, Yingxue Zhang
Abstract
World Action Models (WAMs) couple robot action prediction with video world models. Existing WAMs with shared-backbone and Mixture-of-Transformers designs generally tie the depth of the action module to that of the video backbone, resulting in substantial computational overhead and high inference latency. To address this limitation, we introduce Dock of Transformer (DoT), a video-centric design principle that treats a pretrained video Transformer as a representation hub and connects lightweight output-heads through docking interfaces. This enables flexible output-head design while providing direct access to representations from all layers of the backbone. We then introduce \textbf{Faster-WAM}, an instantiation of DoT for WAMs, which docks a single-layer action head onto a 30-layer video backbone. The docking interface fuses keys and values from all video layers and applies RoPE realignment. Without additional embodied pretraining, Faster-WAM achieves competitive performance on LIBERO and RoboTwin 2.0 while demonstrating strong out-of-distribution generalization on LIBERO-Plus. Faster-WAM also achieves the lowest end-to-end latency in our controlled comparison, requiring only 66.5 ms per inference --- a \(3.2\times\) speedup over Fast-WAM. Overall, these results demonstrate that the video-centric DoT architecture supports flexible task-specific head design while delivering low inference latency, strong action-prediction performance, and robust generalization.