4D-WAM: Training world action models with 3D trajectory supervision, zero inference overhead

量子位 · wechat · 2026-08-16

4D-WAM introduces a model-agnostic training strategy using 3D trajectory fields to teach world action models (WAMs) dynamic 3D understanding via MotionAlignment and DestinationAlignment. Tested on FastWAM-Joint and Lingbot-VA, it improves performance on standard benchmarks and out-of-distribution perturbations, and boosts real robot task success. Training overhead 2%, inference overhead zero.

Original post →

More from Embodied

Embodied channel →