Zhu Jun's team formalizes General World Models: a five-level roadmap and 130k hours of training data
机器之心 · wechat · 2026-09-10
JiQiZhiXin digests Tsinghua's General World Models from First-Principles paper: Understanding-Imagination-Action as a closed loop, a five-level roadmap (L1 generation to L5 world organizer), and a data pyramid from web video to robot trajectories. Motus2 shares parameters across policy, simulator, and evaluator, trained on 130k hours of egocentric video plus 100+ hours of robot data with a tactile expert. MoT architecture balances unified modeling against modality conflicts.
More from Embodied
- SyncWorld: in-context robot world modeling with few visual interactions — du_yilun · 2026-09-10
- Play2Perfect accepted to CoRL 2026 with in-browser zero-shot assembly demo — leto__jean · 2026-09-10
- GenrobotAI's Whole-Body Mesh Data Captures What Egocentric Robot Learning Has Been Missing — CyberRobooo · 2026-09-10
- STM32 Motor Board Running in Under an Hour with Copperkit — debreuil · 2026-09-10
- Maven Robotics raises $100M Series A led by RoboStrategy, emerges from stealth — Rewkang · 2026-09-10
- Unitree Has Sold 18,000 Humanoid Robots; Wang Xingxing Defines Embodied AI's 'ChatGPT Moment' — CyberRobooo · 2026-09-10