BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries

BridgeWAM. Bridging Video and Action Experts via Latent Bridge Queries. By Jiaxin Fang and eight others, from Fudan University, CocoMatrix, and four other institutions. Published on the tenth of October. Its code repository is worked out: about four hundred source files, with a readme and a licence.

World Action Models, or WAMs, typically join a video expert and an action expert through dense Mixture-of-Transformers attention, applied layer by layer. Two costs follow from that coupling. Action prediction is exposed to visual features that are irrelevant to it, which makes selective use of the visual signal hard. And the two experts are forced to have matching depths, which brings redundant computation with it.

The proposal is called Bridge-of-Experts, or BoE, and its instance here is BridgeWAM. Instead of wiring every layer of one expert to every layer of the other, it routes what passes between them through a narrow channel: thirty-two Latent Bridge Queries, or LBQs. These aggregate visual, language, and proprioceptive priors into a shared latent space of limited size.

Training supervises video and action jointly. That lets gradients from the action side steer the Video Expert toward representations better suited to predicting actions. Meanwhile, the Action Expert learns to retrieve task-relevant information selectively, through cross-attention. With the bridge in place, the two experts no longer need the same depth, and the LBQs are reused during action denoising. The result is a model of five point zero six billion parameters, pairing a thirty-layer Video Expert with an Action Expert of only two layers.

That model reaches ninety-eight point two five percent success on LIBERO, and seventy-two point two four percent pooled success on LIBERO-Plus, which the authors present as effective task execution and strong generalization from an ultralight action head. Ablations support two of the design choices: adapting representations jointly, and conditioning on the LBQs in a unified way.

Against Fast-WAM, BridgeWAM cuts single-sample inference FLOPs by twenty-eight point seven two percent in simulation profiling. On the real robot, latency drops to seventy-seven point one milliseconds, a speedup of two point four seven times. Qualitatively, attention favours target objects and task destinations across perturbations.

A two-layer action head on top of a thirty-layer video model sounds lopsided. Read through the abstract, the imbalance is the design: the heavy lifting stays in the Video Expert, and a small, fixed set of queries carries across only what the action side asks for.