PulseAugur
EN
LIVE 08:59:58

New AI architecture HWM improves long-horizon planning with multi-scale world models

Researchers have developed Hierarchical Planning with Latent World Models (HWM), a new architecture for improving long-horizon planning in embodied AI. HWM utilizes world models trained at multiple temporal scales within a shared latent space, allowing predictions from longer-horizon models to guide shorter-horizon planning without explicit task-specific rewards or policies. The system also incorporates an action encoder to compress primitive actions into latent macro-actions, making long-horizon search more tractable. In real-world manipulation tasks, HWM achieved a 70% success rate, a significant improvement over single-level planning which yielded 0% success, and demonstrated up to a 3x reduction in planning compute on simulated tasks. AI

IMPACT This research could significantly improve the capabilities of embodied AI agents in complex, long-horizon tasks.

RANK_REASON This is a research paper detailing a new AI architecture and planning paradigm. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

COVERAGE [1]

  1. arXiv cs.LG TIER_1 English(EN) · Wancong Zhang, Basile Terver, Artem Zholus, Soham Chitnis, Harsh Sutaria, Mido Assran, Randall Balestriero, Amir Bar, Adrien Bardes, Yann LeCun, Nicolas Ballas ·

    Hierarchical Planning with Latent World Models

    arXiv:2604.03208v2 Announce Type: replace Abstract: World models are a promising path to zero-shot embodied control through planning. However, existing world model planners struggle on long-horizon, multi-stage tasks: prediction errors compound and naive search is exponential in …