Faster-WAM: Do World Action Models Need Deep Action Modules?

Existing World Action Models (WAMs) often use a Mixture-of-Transformers (MoT) design that ties the depth of the action module to the video backbone. This creates significant computational overhead and high inference latency. The authors ask: do WAMs truly require deep, mirrored action modules, or can a lightweight, flexible head access the rich physical dynamics already captured by a pretrained video backbone?

The authors introduce the Dock of Transformer (DoT) design principle. Instead of treating the action module as a peer to the video backbone, DoT treats the video backbone as a central 'representation hub.' Lightweight task-specific heads are attached via a 'docking interface' that aggregates keys and values from all layers of the backbone.

Faster-WAM is the specific instantiation of this principle. It uses a 30-layer video backbone and docks a single-layer action head onto it. The docking interface uses two key mechanisms:

1. KV-Fusion: Aggregates and remaps key-value representations from all video layers into the action head's feature space.
2. Video-Action RoPE Alignment: Corrects the mismatch between the 3D rotary positional embeddings (RoPE) used in the video backbone and the 1D RoPE used in the action head, ensuring the action head can correctly interpret the spatial-temporal representations.

Faster-WAM achieves a 3.2x speedup over the previous state-of-the-art (Fast-WAM), reducing inference latency to 66.5 ms per action chunk. Despite the drastic reduction in action-head depth, the model maintains high performance on standard benchmarks, achieving 98.5% success on LIBERO and 89.17% on RoboTwin 2.0. It also demonstrates strong out-of-distribution generalization on the LIBERO-Plus suite, outperforming previous models without requiring additional embodied pretraining.

This work challenges the assumption that action prediction in robotics requires deep, complex architectures that mirror the visual backbone. By demonstrating that a lightweight, docked head can effectively leverage pre-existing representations from a video world model, the authors provide a pathway to building high-performance, low-latency robotic controllers that are more suitable for real-time deployment.