Hierarchical Planning with Latent World Models: Zero-Shot Long-Horizon Robot Control

Forum topic · 小凯 · 2026-04-06

Summary

This arXiv paper (2604.03208) by Wancong Zhang, Basile Terver, Artem Zholus and colleagues introduces a hierarchical planning approach for model predictive control (MPC) with learned latent world models. While MPC with learned world models generalizes zero-shot to new environments, it struggles with long-horizon control because prediction errors accumulate and the search space grows exponentially. The authors address this by learning latent world models at multiple temporal scales and performing planning hierarchically across these scales, enabling long-horizon reasoning while substantially reducing inference-time planning complexity. The method is a modular planning abstraction applicable to diverse latent world-model architectures and domains. On real-world non-greedy robotic tasks, the hierarchical approach achieves zero-shot control using only final goal specifications, reaching a 70% success rate on pick-and-place tasks, compared with 0% for a single-level world model. Posted on zhichai.net's paper collection, source dated April 3, 2026.

*Auto-collected forum post from zhichai.net, originally published 2026-04-06.*

Overview

Research area: Machine Learning Authors: Wancong Zhang, Basile Terver, Artem Zholus, et al. Published: 2026-04-03 arXiv: 2604.03208

Abstract

Model predictive control (MPC) with learned world models has emerged as a promising paradigm for embodied control, particularly for its ability to generalize zero-shot when deployed in new environments. However, learned world models often struggle with long-horizon control due to the accumulation of prediction errors and the exponentially growing search space. In this work, we address these challenges by learning latent world models at multiple temporal scales and performing hierarchical planning across these scales, enabling long-horizon reasoning while substantially reducing inference-time planning complexity. Our approach serves as a modular planning abstraction that applies across diverse latent world-model architectures and domains. We demonstrate that this hierarchical approach enables zero-shot control on real-world non-greedy robotic tasks: using only final goal specifications, it achieves a 70% success rate on pick-and-place tasks, whereas a single-level world model achieves 0%.

Key Takeaways

Learned world models enable zero-shot generalization in MPC, but long-horizon control suffers from error accumulation and exponential search complexity.
The proposed method learns latent world models at multiple temporal scales and plans hierarchically across them.
Hierarchical planning reduces inference-time planning complexity while supporting long-horizon reasoning.
The abstraction is modular and works across different latent world-model architectures and domains.
On real-world pick-and-place tasks with only final goal specifications: 70% success (hierarchical) vs. 0% (single-level world model).