BridgeWAM Decouples Video and Action Experts via Latent Bridge Queries
October 9, 2026
BridgeWAM replaces dense, layer-wise Mixture-of-Transformers with 32 Latent Bridge Queries to aggregate visual, language, and proprioceptive priors. This architecture decouples expert depths and uses joint supervision to allow action gradients to guide video representations toward task-relevant features.
HOW THIS AFFECTS YOU
●
builderThis method may reduce redundant computation in robot action prediction pipelines.
●
researcherYou can explore decoupled expert architectures for more efficient multimodal learning.