RL-VLA 3 : Reinforcement Learning VLA Accelerating via Full Asynchronism

Zhong Guan, Haoran Sun, Yongjian Guo, Shuai Di, Xiaodong Bai, Jing Long, Tianyun Zhao, Mingxi Luo, Chen Zhou, Yucheng Guo, Qiming Yang, Wanting Xu, Wen Huang, Yunxuan Ma, Hongke Zhao, Likang Wu, Xiaotie Deng, Xi Xiao, Sheng Wen, Yicheng Gong, Junwu Xiong · 2026

Read RL-VLA 3 : Reinforcement Learning VLA Accelerating via Full Asynchronism by Zhong Guan, Haoran Sun, Yongjian Guo, Shuai Di, Xiaodong Bai, Jing Long, Tianyun Zhao, Mingxi Luo, Chen Zhou, Yucheng Guo, Qiming Yang, Wanting Xu, Wen Huang, Yunxuan Ma, Hongke Zhao, Likang Wu, Xiaotie Deng, Xi Xiao, Sheng Wen, Yicheng Gong, Junwu Xiong (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-05765.txt.

Description

Robot policies that learn from fake worlds suffer when training waits on slow, uneven physics engines. Haoran Sun and 15 co-authors build an out-of-step reward learning frame for vision language action models, the policies that drive embodied agents. Standard frames switch strictly between runs and tuning, treating each run as one whole unit, a design taken from language model training.

Their system lets fake-world runs, guessing, and training run at once, with smart batch planners and flexible world splitting. Tests span several fake-world backends, model shapes, and learning rules. Throughput rose up to eighty five percent over lockstep rivals while sample use stayed the same. Scaling held from a handful of graphics chips to hundreds. The authors describe the system as the first fully out-of-step frame built for vision language action training. It targets robot teams whose choke point is setup, not rules.

Catalog record

CREATORZhong Guan, Haoran Sun, Yongjian Guo, Shuai Di, Xiaodong Bai, Jing Long, Tianyun Zhao, Mingxi Luo, Chen Zhou, Yucheng Guo, Qiming Yang, Wanting Xu, Wen Huang, Yunxuan Ma, Hongke Zhao, Likang Wu, Xiaotie Deng, Xi Xiao, Sheng Wen, Yicheng Gong, Junwu Xiong
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE36 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-05765

Browse the full library · About the Tower of Babel