4D-WAM: Infusing Spatiotemporal Awareness into World Action Models via Trajectory Fields
zhenjun_zhao · x · 2026-08-13
World-Action Models (WAMs) typically represent videos in 2D pixel space, creating a representation gap with the 3D space where robotic actions are executed. Existing 3D approaches fail to fully exploit the dynamics of 3D structures. This paper proposes 4D-WAM, a model-agnostic training strategy.
- Core Method: Injects spatiotemporal knowledge from 3D trajectory fields into WAMs through representation alignment.
- Dual Objectives: 1) Motion alignment aligns temporal feature variations across adjacent frames to build local 4D awareness; 2) Destination alignment guides the model to infer the final destination by minimizing the gap between attention-like similarity distributions of the source and destination frames.
- Results: These objectives provide local motion supervision and long-horizon goal guidance, enabling WAMs to learn trajectory-level spatiotemporal representations, performing well in in-distribution and out-of-distribution experiments.
More from Embodied
- DynaRobotics CEO Lindon Gao on running deployment and research feedback loops together — micoolcho · 2026-10-16
- AutoAgent0: VLM-powered agentic runtime lifts long-tail driving success by 28.3pp — yining_hong · 2026-10-16
- CarryOn: dual-loop harness adds personalization and safe recovery to frozen VLA robots — Zixuan Wang · 2026-10-16
- FoveaVLA replaces wrist cameras with learned foveation, hitting 70.8% success — Haoyu_Xiong_ · 2026-10-16
- Musk teases two-way Grok x Optimus integration: Grok in the bot, bot in Grok — beffjezos · 2026-10-16
- SimVTA uses simulator-rendered video to post-train driving world action models with RL — OpenDriveLab · 2026-10-16