BridgeWAM: latent bridge queries decouple video and action experts, 2.47x faster robot inference
Fudan-University · hf · 2026-10-19
Fudan, Tsinghua, and Peking University researchers propose BridgeWAM, a World Action Model using Bridge-of-Experts design. Instead of dense layer-wise MoT attention coupling video and action experts, 32 shared latent bridge queries (LBQs) aggregate vision, language, and proprioception priors, decoupling the two experts: action gradients steer the video expert's representations while the action expert retrieves task-relevant features via cross-attention.
The 5.06B model (30-layer video expert, 2-layer action expert) achieves 98.25% success on LIBERO and 72.24% pooled success on LIBERO-Plus. Versus Fast-WAM it cuts single-sample inference FLOPs by 28.72% and real-robot latency to 77.1ms (2.47x speedup).
More from Embodied
- Asimov open-sources Python SDK and Edge control layer for its humanoid robot — freelerobot · 2026-10-19
- Tesla Roadster Stunts Simulated in MuJoCo with Thruster and Adhesion Actuators — kevin_zakka · 2026-10-19
- MorphAct: coupled low-rank generators dynamically adapt robot policies across contexts — SJTU · 2026-10-19
- Dev builds a hand-tracking game in hours with Tesla Specs glasses' CLAD — stspanho · 2026-10-19
- Does a robot's jump count as flight? X users pedantically debate the definition — basedjensen · 2026-10-19
- MolmoAct2 Lead Duan Jiafei Joins NUS, Wins Multimillion-Dollar Embodied AI Grant — DJiafei · 2026-10-19