Fac-TDMPC: A Factored World Model for Robot Planning

TMLR Paper6068 Authors

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 suffer 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 \textbf{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.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Florian_Shkurti1

Submission Number: 6068