Publications

• Sorted by Date • Classified by Publication Type • Classified by Research Category •

Fac-TDMPC: A Factored World Model for Robot Planning

Yuan Zhang, Jianhong Wang, Jinke He, Frans A Oliehoek, and Joschka Boedecker. Fac-TDMPC: A Factored World Model for Robot Planning. Transactions on Machine Learning Research, 2026.

Download

pdf [12.2MB]  

Abstract

Model-based reinforcement learning (MBRL) has shown strong sample efficiency in robotics by learning predictive world models and planning with them, but existing methods suf- fer from high planning latency due to the combination of centralized world models and model predictive control (MPC) as planners, thus limiting the real-time deployment in high-dimensional action spaces. We introduce Fac-TDMPC, a factored latent-space world model that decomposes transition, reward, and value functions on the latent space and learns the factorization via model distillation. The factored design enables decentralized planning across action dimensions. Empirically, Fac-TDMPC achieves substantial planning speedups while preserving the control performance across a suite of continuous-control robotic tasks; it also demonstrates improved robustness to action perturbations, interpretable joint-level latent structure, and enhanced multi-task data efficiency.

BibTeX Entry

@article{Zhang26TMLR,
    title=      {Fac-{TDMPC}: A Factored World Model for Robot Planning},
    author=     {Yuan Zhang and Jianhong Wang and Jinke He and Frans A Oliehoek and Joschka Boedecker},
    journal=    TMLR,
    issn=       {2835-8856},
    year=       {2026},
    OPTurl=        {https://openreview.net/forum?id=Smb0sdocmd},
    note=       {}
    keywords =  {refereed},
    abstract  = {
    Model-based reinforcement learning (MBRL) has shown strong sample efficiency in robotics
    by learning predictive world models and planning with them, but existing methods suf-
    fer from high planning latency due to the combination of centralized world models and
    model predictive control (MPC) as planners, thus limiting the real-time deployment in
    high-dimensional action spaces. We introduce Fac-TDMPC, a factored latent-space world
    model that decomposes transition, reward, and value functions on the latent space and learns
    the factorization via model distillation. The factored design enables decentralized planning
    across action dimensions. Empirically, Fac-TDMPC achieves substantial planning speedups
    while preserving the control performance across a suite of continuous-control robotic tasks;
    it also demonstrates improved robustness to action perturbations, interpretable joint-level
    latent structure, and enhanced multi-task data efficiency.
    }
}

Generated by bib2html.pl (written by Patrick Riley) on Thu Oct 08, 2026 13:51:42 UTC