Driving · 2026-10-14

BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries

Jiaxin Fang, Ruitong Liu, Zeyu Lai, Xingyu Wang, Mingyuan Jia, Pingyue Sheng, Weikang Wan, Mengdi Xu, Yuxiang Gao

arXiv:2610.14746PDF

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

BridgeWAM (Bridge-of-Experts) uses 32 Latent Bridge Queries (LBQs) to connect a 30-layer Video Expert with a two-layer Action Expert in a shared limited latent space. Joint video–action supervision routes action gradients through the LBQs to adapt video representations for control, while the Action Expert selectively retrieves from the LBQ latent via cross-attention. The authors report 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus.

Why it matters

The authors address limitations of layer-wise MoT coupling in world action models: dense action readout from video can attend to irrelevant visual features, and one-to-one layer depth matching can add redundant computation and parameters. BridgeWAM proposes a compact latent interface (LBQs) that decouples expert depths and reuses a single LBQ conditioning during action denoising, targeting both better selective feature use and lower inference latency.

Method

  • Introduce Bridge-of-Experts (BoE) with 32 Latent Bridge Queries (LBQs) that aggregate visual, language, and proprioceptive priors into a shared limited latent space for action prediction.
  • Use a 30-layer Video Expert with LBQs that are computed once from the current context and reused throughout action denoising; action gradients flow through the LBQ pathway to guide video representation learning.
  • Decouple expert depths with a two-layer alternating cross–self-attention Action Expert that reads information from LBQs for action-chunk coordination.

Limitation

The authors report low scores for some LIBERO-Pro perturbation groups (swap and task), indicating that substantial limitations can be obscured by aggregate results.

Abstract (from arXiv)

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. We introduce Bridge-of-Experts (BoE), instantiated as BridgeWAM, which aggregates visual, language, and proprioceptive priors through 32 Latent Bridge Queries (LBQs) into a shared, limited latent space. Joint video-action supervision lets action gradients guide the Video Expert toward representations better suited for action prediction, while the Action Expert learns to selectively retrieve task-relevant information through cross-attention. Decoupling expert depths and reusing LBQs during action denoising yield a 5.06B model with a 30-layer Video Expert and a two-layer Action Expert, achieving 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus, demonstrating effective task execution and strong generalization with an ultralight action head. Ablations support joint representation adaptation and unified LBQ conditioning. Relative to Fast-WAM, BridgeWAM reduces single-sample inference FLOPs by 28.72% in simulation profiling and real-robot latency to 77.1 ms (a 2.47-fold speedup). Qualitative experimental results further indicate preferential attention to target objects and task destinations across perturbations.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-10 · Indexed by Pub-AI: 2026-10-14 UTC

preprint

BridgeWAM: Bridging Video and Action Experts via Latent Bridge Queries

Jiaxin Fang, Ruitong Liu, Zeyu Lai, Xingyu Wang, Mingyuan Jia, Pingyue Sheng, Weikang Wan, Mengdi Xu, Yuxiang Gao

Preprint posted: 2026-10-10 · Original record ↗

arxiv
2610.14746
Explicit source versions (1)
  • v1 · 2026-10-10
  1. Preprint available

    Source date: 2026-10-10 · First detected: 2026-10-25 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

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. We introduce Bridge-of-Experts (BoE), instantiated as BridgeWAM, which aggregates visual, language, and proprioceptive priors through 32 Latent Bridge Queries (LBQs) into a shared, limited latent space. Joint video-action supervision lets action gradients guide the Video Expert toward representations better suited for action prediction, while the Action Expert learns to selectively retrieve task-relevant information through cross-attention. Decoupling expert depths and reusing LBQs during action denoising yield a 5.06B model with a 30-layer Video Expert and a two-layer Action Expert, achieving 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus, demonstrating effective task execution and strong generalization with an ultralight action head. Ablations support joint representation adaptation and unified LBQ conditioning. Relative to Fast-WAM, BridgeWAM reduces single-sample inference FLOPs by 28.72% in simulation profiling and real-robot latency to 77.1 ms (a 2.47-fold speedup). Qualitative experimental results further indicate preferential attention to target objects and task destinations across perturbations.