Implicit Worldline Constraint Models: Joint Trajectory Energy Optimization Without Autoregressive Rollout
Year: 2026
Type: preprint
Abstract: Learned world models based on autoregressive rollout suffer from compounding prediction errors that accumulate over long planning horizons. This paper introduces Implicit Worldline Constraint Models (IWCM), a constraint-based world modeling framework that represents future trajectories through joint optimization problems over latent worldlines rather than sequential predictions. Instead of generating states autoregressively, IWCM learns a differentiable energy function over complete trajectories and performs planning through gradient-based constraint satisfaction. The paper presents experiments on synthetic grid-world environments and continuous-control tasks from DM Control, including comparisons with autoregressive rollout baselines, initialization ablations, recovery from degraded rollouts, and a self-supervised pixel pipeline (TAMG) that replaces oracle state representations with learned object-centric slot representations. Code, training pipelines, and reproduction scripts are provided to facilitate reproducibility.
Author: Pérez Muñiz, Zabdiel
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7166541599
Cites:
Cited by:
Related to: