A new research paper, "Every Camera Grows Up: Camera-Height Regression in Interactive Video World Models," highlights a significant flaw in six recent open-weight models. These models, designed to simulate embodied agents, incorrectly regress camera viewpoints towards human eye level (1.1m to 1.5m) instead of maintaining the initial camera height. This "camera-height regression" occurs even when the starting view is still present and worsens with increased rotation, impacting visual navigation policies. AI
IMPACT This regression flaw could impact the reliability of AI simulators for embodied agents and visual navigation.
RANK_REASON The cluster contains a research paper detailing a novel finding about AI model behavior. [lever_c_demoted from research: ic=1 ai=1.0]
- Every Camera Grows Up: Camera-Height Regression in Interactive Video World Models
- Matrix-Game
- SANA-WM
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →