Robots learn contact-rich manipulation in world models without simulators

Joseph Amigo et al. (NYU/ISIR) released a preprint coupling local latent models with global world models so robots can learn contact-rich manipulation entirely without a simulator, deploying zero-shot on real hardware.

2026-09-04 ~ 2026-09-04 · 2 related posts

• Training contact-rich humanoid manipulation policies purely inside a world model — chris_j_paxton · 2026-09-04
• Policies trained fully inside world models deploy zero-shot on real robots, beating PPO — chris_j_paxton · 2026-09-04