機器人與世界模型
Faster-WAM Shrinks the Robot Action Head to a Single Layer, Cutting Control Inference Latency to 66.5 Milliseconds
Faster-WAM attaches a single-layer action head to a 30-layer video Transformer and obtains world-model representations through cross-layer KV fusion, eliminating the need for the action branch to replicate the full backbone depth. The study reports a 3.2× speedup over Fast-WAM, but code and weights for the new architecture have not yet been released, and the results come primarily from simulation benchmarks.
World Action Models (WAMs) jointly learn future visual changes and robot actions, but common shared-backbone or Mixture-of-Transformers designs make the action module as deep as the video model. Faster-WAM, submitted on August 3, challenges this symmetry assumption: if the video Transformer already forms sufficient spatial, semantic, and dynamics representations across its layers, the action side may only need to read those representations instead of repeating dozens of layers of computation.
The team proposes a Dock of Transformer (DoT) interface that treats the pretrained video Transformer as a representation hub. Faster-WAM uses a 30-layer video backbone but only a single-layer action head. The interface collects the keys and values from every video layer, fuses them, realigns their positions using RoPE, and passes them to the action head, allowing the shallow head to access visual information at different levels of abstraction. This complements Fast-WAM’s earlier approach of removing explicit future-video generation during inference: video modeling primarily provides representations during training, while online control produces actions directly.
Under identical control conditions, the paper reports an inference latency of 66.5 milliseconds per run, 3.2× faster than the controlled Fast-WAM baseline. Without additional embodied pretraining, the model maintains competitive success rates on LIBERO and RoboTwin 2.0, while LIBERO-Plus evaluates its generalization under shifts in appearance, objects, and instruction distributions. These findings suggest that WAM latency bottlenecks need not be addressed solely through quantization, distillation, or fewer denoising steps; the depth assigned to each modality branch is itself an architectural parameter that can be optimized independently.
However, the 66.5-millisecond figure is an end-to-end measurement from the paper’s controlled environment and should not be treated as equivalent to the sensing, communication, and actuation loop of a physical robot. No public weights or official implementation of the new Faster-WAM architecture have been found so far; the existing Fast-WAM repository can only help clarify the predecessor’s training and evaluation setup. Key issues to watch include the memory cost of cross-layer KV fusion, portability across different video backbones, and whether a shallow action head sacrifices contact-control performance and error-recovery capabilities during long-horizon physical manipulation.