# The training objective that makes robot videos look great but fail at contact

- slug: the-training-objective-that-makes-robot-videos-look-great-but-fail-at-contact
- date: 2026-03-23
- category: Robotics

Researchers from Zhejiang University, Westlake, and Beijing AI propose VAMPO, a reinforcement learning approach to fix the 'looks right, fails anyway' problem in diffusion-based video action models.

---