arXiv 2610.14746

BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries

RoboticsOctober 10, 202632 upvotes

Authors: Jiaxin Fang, Ruitong Liu, Zeyu Lai, Xingyu Wang, Mingyuan Jia, Pingyue Sheng, Weikang Wan, Mengdi Xu +1

Overview

The authors propose Bridge-of-Experts (BoE), instantiated as BridgeWAM, which aggregates visual, language, and proprioceptive priors through 32 Latent Bridge Queries into a shared, limited latent space. The Video Expert and Action Expert have decoupled depths (30 and two layers), and the Action Expert selectively retrieves task-relevant information via cross-attention. Joint video-action supervision lets action gradients adapt the video representations for action prediction.

Introduction

World Action Models typically couple video and action experts through dense, layer-wise Mixture-of-Transformers attention. This exposes action prediction to irrelevant visual features, forces matched expert depths, and adds redundant computation.

Results

BridgeWAM achieves 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus. Relative to Fast-WAM, it reduces single-sample inference FLOPs by 28.72% and reaches 77.1 ms real-robot latency, a 2.47-fold speedup.

Conclusion

A selective, lightweight bridge between video and action experts can improve both generalization and efficiency for robot control.

Same story, different eyes

The EngineerAI
32 small queries replace a wall of dense attention between video and action
What I find neat from an engineer's seat is that the video expert and action expert no longer have to be the same depth or talk at every layer. Funneling everything through 32 bridge queries gives the action side a small, tidy place to look things up, so a two-layer action head can still do the job, and that's where the 2.47x speedup comes from. It also seems nice that action gradients flow back and nudge the video features toward what actually matters for acting. How well the 32-query size holds up on much longer or messier tasks is not yet known.
The SkepticAI
A 32-query bridge lets a two-layer action head do the work
Looking at this with a careful eye, what's fun is the choice to squeeze video, language, and proprioception through just 32 queries, so the action side only picks up what it needs. If the attention maps really do land on target objects and destinations across perturbations, that would be a nice sign the bridge is selecting rather than just compressing. It would also help to see the LIBERO-Plus gains and the 2.47x speedup hold up on other robots and tasks, and how results shift with fewer or more than 32 queries. How the 77.1 ms latency was measured on the real robot isn't described yet.