MECo-WAM: 4D geometric priors boost robot manipulation to 98.2%

World Action Models (WAMs) jointly model visual dynamics and action sequences for robotic manipulation, but existing methods optimize appearance-oriented latents that miss crucial evolving geometry. Researchers from multiple institutions propose MECo-WAM (Multi-Expert Co-Training World Action Model) to inject action-relevant 4D geometric priors directly into video-action representations while preserving the lightweight inference graph. During training, MECo-WAM combines video and action experts with a 4D expert supervised by relational targets from a frozen VGGT encoder. Asymmetric expert visibility prevents non-causal shortcuts from auxiliary geometry to action generation.

To transfer geometric knowledge into the deployed pathway, the team introduces decayed 4D read-mask attention—providing restricted current-frame geometric guidance early in training and progressively removing this dependency. They also propose action-aware temporal geometric distillation, which aligns within-frame geometric relations and temporal evolution while emphasizing visual regions most relevant to robot actions. At deployment, all auxiliary 4D components are removed, yielding no inference cost increase. Experiments on LIBERO (98.2%), RoboTwin 2.0 (92.6%), and real-world tasks confirm that MECo-WAM improves manipulation performance without added compute overhead—a significant step for practical robotics.