Model family

Z-sight

Z-sight is a 3.18-billion-parameter research robot policy that combines vision-language-action modeling with latent world modeling. It jointly denoises future visual tokens and continuous actions, and was evaluated in simulation and on a physical Unitree G1.

Overview

Model type
Action policy
Architecture
VLA · WAM · Transformer
Action generation
Flow matching
Listed versions
Z-sight
Related robots
1

Papers & reports (1)

Z-sight: Unifying Vision-Language-Action and Latent World Modeling

Paper available

arXiv: 2611.24940DOI: 10.48550/arxiv.2611.24940
Authors & affiliations (15)

Author order: Original order

  1. Junjie Ye

    Maven Robotics · University of Southern California

  2. Yiqi Zhao

    University of Southern California

  3. Muchen Xu

    University of Southern California

  4. Celina Shiyu Wang

    University of Southern California

  5. Hao Dong

    University of Southern California

  6. Ruohai Ge

    University of Southern California

  7. Rong Xue

    University of Southern California

  8. Benjamin J Kolligs

    Maven Robotics

  9. Aditya Dutt

    Maven Robotics

First published
2026-11-19
Latest paper revision
v1
Venue
Not provided

Covered model versions

Paper ↗PDF ↗

Official resources

Submit a source or correction