VFlash jointly predicts visual primitive coordinates
An arXiv paper introduces VFlash, a visual-planning method that uses Flash Primitive Decoding to predict a primitive’s coordinates jointly rather than one token at a time. Across evaluated 2D and 3D robot trajectory tasks, the method reportedly reduced prediction errors relative to autoregressive decoding. Preliminary single-rollout measurements showed approximately 25x faster generation for 3D point motion forecasting. A world action model conditioned on VFlash plans achieved a reported 70.2% average success rate across four MolmoSpaces task suites, compared with 62.6% for the baseline.
Visual plans represent spatial goals and anticipated motion in robotics and autonomous driving. The paper reports that generating these plans one coordinate token at a time makes richer predictions increasingly costly, while VFlash seeks to reduce prediction errors and generation time.
Computer Science > Robotics
arXiv:2611.08449 (cs)
[Submitted on 5 Nov 2026]
Title:VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
Authors:Yu-Cheng Chou, Shihao Wang, Jin Huang, Chi-Pin Huang, Chiao-An Yang, Xianzhe Fan, Jing Wang, Zhiqi Li, Yuanguo Kuang, Ruijie Zheng, Johan Bjorck, Sifei Liu, Hang Su, Abhishek Badki, Lei Zhang, Alan Yuille, Linxi "Jim'' Fan, Jan Kautz, Zhiding Yu
View a PDF of the paper titled VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding, by Yu-Cheng Chou and 18 other authors