🛰️ Daily AI Frontier
‹ back to 2026-08-29

4DSynth: Controllable Procedural World Synthesis for Dynamic Embodied Simulation

arXiv cs.RO Multimodal & Generative Zehao Qi, Haochen Luo, Jia-Wang Bian, Zeyu Ma, Shuyang Sun 2026-08-27
Representative image for 4DSynth: Controllable Procedural World Synthesis for Dynamic Embodied Simulation

TL;DR - 4DSynth generates editable, physics-ready 4D simulation environments from text, blueprint masks, or single images. It also enables reproducible, tunable benchmarks for developing and evaluating embodied agents.

  • Produces explicit geometry, animated actors, collision-free trajectories, and simulation-ready physical state.
  • Uses a shared geometry-grounded representation for animation, camera planning, rendering, and task generation.
  • Introduces 4DSynth-Nav, an interactive navigation benchmark generated entirely from procedural scenes.
  • Two vision-language models failed most navigation tasks across three difficulty tiers, often stalling after early subtasks.

view merged work →