Z-Sight Unifies Vision-Language-Action and Latent World Modeling for Robotics

Z-Sight Unifies Vision-Language-Action and Latent World Modeling for Robotics

Junjie Ye, Yiqi Zhao, Muchen Xu, Celina Shiyu Wang, Hao Dong +10 more

2 min read•Nov 19, 2026

Conclusion

Limitations

The foresight stream currently predicts the future from the observation alone. Conditioning it on candidate actions would let Z-sight compare the outcomes of different actions and open the door to test-time planning in its own representation space.

Hardware and Sonic Teleoperation-Deployment Pipeline

Hardware and teleoperation-deployment pipeline

We deploy real-world policies on a Unitree G1 humanoid equipped with two 7-DoF Dex3-1 dexterous hands and a single egocentric head camera. The default head-mounted Intel RealSense D435i provides a narrow field of view suited to tabletop tasks (Dog and Puffs), whereas tasks involving locomotion (Chair, Book, Trash, and Fruit) require wider peripheral visibility and use a head-mounted ZED 2i camera. Each task uses the same camera during data collection and evaluation.

Humanoid robot and egocentric camera task setup

Frequently Asked Questions

What does the foresight stream currently predict from? It predicts the future from the observation alone.

How could candidate actions change the foresight stream? Conditioning on candidate actions could let Z-sight compare their outcomes and enable test-time planning in its own representation space.

Which humanoid is used for real-world policy deployment? The policies are deployed on a Unitree G1 humanoid.

Which cameras are used for the tasks? Tabletop tasks use a head-mounted Intel RealSense D435i, while locomotion tasks use a head-mounted ZED 2i. Each task uses the same camera for data collection and evaluation.