BridgeWAM links video and action experts with latent queries
Original title: BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries
The 5.06B model reports 98.25% LIBERO success and 28.72% fewer inference FLOPs than Fast-WAM.
BridgeWAM aggregates visual, language and proprioceptive priors through 32 latent bridge queries instead of dense layer-wise mixture-of-transformers attention. Joint supervision steers the video expert, while the action expert retrieves task-relevant features by cross-attention. The paper reports 72.24% pooled LIBERO-Plus success and 77.1 ms real-robot latency, a 2.47-fold speedup versus Fast-WAM.
- The video expert has 30 layers and the action expert has two.
- Simulation profiling shows 28.72% fewer single-sample FLOPs than Fast-WAM.
- Ablations support joint representation adaptation and unified query conditioning.
From the original article
1 min read at huggingface.co
World Action Models (WAMs) commonly couple video and action experts through dense, layer-wise Mixture-of-Transformers (MoT) attention, exposing action prediction to irrelevant visual features while enforcing matched expert depths. This coupling hinders selective visual-feature use and introduces redundant computation.
By Jiaxin Fang, Ruitong Liu, Zeyu Lai · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.