Contact-Rich Humanoid Policies Trained Entirely Inside a World Model

Loading video
Loading videoCoupled Local and Global World Models uses a large-scale world model for high-fidelity rollouts and a lightweight latent surrogate for gradients, training contact-rich G1 humanoid policies entirely inside a learned pixel-space model before zero-shot hardware deployment.
Coupled Local and Global World ModelsFoGFirst-Order RLDMOWorld ModelsUnitree G1UnitreeHumanoidManipulationZero-ShotReinforcement LearningRoboPapersContact-Rich Manipulation
Category: research
Author: @chris_j_paxton
Date: 2026-09-04T00:00:00
Duration: 145.833s
Reference: https://coupled-global-local-wm-rl.pages.dev





