Training contact-rich humanoid manipulation policies purely inside a world model
chris_j_paxton · x · 2026-09-04
- RL has delivered gains in whole-body humanoid control but struggled with manipulation, where capturing contact dynamics in simulation is hard.
- Training inside learned world models is a promising fix, yet powerful world models are too computationally expensive to train in directly.
- A new paper from @Jsphamigo and @Rk4342R decomposes the problem: a large global world model generates forward trajectories, while a lightweight low-dimensional latent-space model approximates local contact dynamics — making contact-rich humanoid manipulation training tractable purely within a world model.
Related event: Robots learn contact-rich manipulation in world models without simulators(2 posts)→
More from Embodied
- Beni becomes most-funded personal robot on Kickstarter, mass production underway for Sept-Oct delivery — ZeYanjie · 2026-09-04
- In-Context Learning Hints at a 'GPT Moment' for Robotics: Demos as the New Interface — chris_j_paxton · 2026-09-04
- Parent turns kid's nightly AI voice journal into private childhood wiki with DayRing for Kids — churchkey · 2026-09-04
- ACE-Data-0: 150 hours of open household robot data with touch, audio and multi-view — CyberRobooo · 2026-09-04
- Unity touts GPT-6 Astra integration for coding agents building 3D environments — chongdashu · 2026-09-04
- Lingverse iKairos is a desktop robot companion that doubles as a neck-worn AI pendant — francisybea · 2026-09-04