Energy-Structured Latent World Models with Neural Time Fields for Physically Constistent Open-World Motion Planning
The paper’s claim is that putting energy and momentum directly into a latent world model makes robot motion forecasts and navigation safer.
The authors propose ELWM, a latent model whose state explicitly carries physical quantities and uses causal transitions through dissipation and control ports. They pair it with Physics-Conditioned Neural Time Fields to turn those predictions into a navigation policy. In held-out scenes, the reported gains include lower 0.8-second prediction error, higher navigation success, better SPL, and a collision-rate drop from 12.1% to 5.8%. ArXiv · AI/CL/LG's note
The authors propose ELWM, a latent model whose state explicitly carries physical quantities and uses causal transitions through dissipation and control ports. They pair it with Physics-Conditioned Neural Time Fields to turn those predictions into a navigation policy. In held-out scenes, the reported gains include lower 0.8-second prediction error, higher navigation success, better SPL, and a collision-rate drop from 12.1% to 5.8%. ArXiv · AI/CL/LG's note
score 4