_version_ 1866914328687411200
author Cai, Junhao
Cai, Zetao
Cao, Jiafei
Chen, Yilun
He, Zeyu
Jiang, Lei
Li, Hang
Li, Hengjie
Li, Yang
Liu, Yufei
Lu, Yanan
Lv, Qi
Ma, Haoxiang
Pang, Jiangmiao
Qiao, Yu
Qiu, Zherui
Shen, Yanqing
Shi, Xu
Tian, Yang
Wang, Bolun
Wang, Hanqing
Wang, Jiaheng
Wang, Tai
Wei, Xueyuan
Wu, Chao
Xie, Yiman
Xing, Boyang
Yang, Yuqiang
Yang, Yuyin
Yu, Qiaojun
Yuan, Feng
Zeng, Jia
Zhang, Jingjing
Zhang, Shenghan
Zhang, Shi
Zhaxi, Zhuoma
Zhou, Bowen
Zhou, Yuanzhen
Zhou, Yunsong
Zhu, Hongrui
Zhu, Yangkun
Zhu, Yuchen
author_facet Cai, Junhao
Cai, Zetao
Cao, Jiafei
Chen, Yilun
He, Zeyu
Jiang, Lei
Li, Hang
Li, Hengjie
Li, Yang
Liu, Yufei
Lu, Yanan
Lv, Qi
Ma, Haoxiang
Pang, Jiangmiao
Qiao, Yu
Qiu, Zherui
Shen, Yanqing
Shi, Xu
Tian, Yang
Wang, Bolun
Wang, Hanqing
Wang, Jiaheng
Wang, Tai
Wei, Xueyuan
Wu, Chao
Xie, Yiman
Xing, Boyang
Yang, Yuqiang
Yang, Yuyin
Yu, Qiaojun
Yuan, Feng
Zeng, Jia
Zhang, Jingjing
Zhang, Shenghan
Zhang, Shi
Zhaxi, Zhuoma
Zhou, Bowen
Zhou, Yuanzhen
Zhou, Yunsong
Zhu, Hongrui
Zhu, Yangkun
Zhu, Yuchen
contents Prevalent Vision-Language-Action (VLA) models are typically built upon Multimodal Large Language Models (MLLMs) and demonstrate exceptional proficiency in semantic understanding, but they inherently lack the capability to deduce physical world dynamics. Consequently, recent approaches have shifted toward World Models, typically formulated via video prediction; however, these methods often suffer from a lack of semantic grounding and exhibit brittleness in the presence of video prediction errors. To synergize semantic understanding with dynamic predictive capabilities, we present InternVLA-A1. This model employs a unified Mixture-of-Transformers architecture, coordinating three experts for scene understanding, visual foresight generation, and action execution. These components interact seamlessly through a unified masked self attention mechanism. Building upon InternVL3 and Qwen3-VL, we instantiate InternVLA-A1 at 2B and 3B parameter scales. We pre-train these models on heterogeneous data sources over real-world robot data, synthetic simulation data, and human videos, covering over 692M frames. This hybrid training strategy effectively harnesses the diversity of synthetic simulation data while minimizing the sim-to-real gap. We evaluated InternVLA-A1 on 12 real-world robotic tasks and a simulation benchmark. The results show that InternVLA-A1 consistently outperforms prior leading models: compared with pi0.5, it achieves +4.4\% on static manipulation tasks and +2.6\% on the RoboTwin 2.0 simulation benchmark, and delivers a +26.7\% boost on dynamic manipulation tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2601_02456
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation
Cai, Junhao
Cai, Zetao
Cao, Jiafei
Chen, Yilun
He, Zeyu
Jiang, Lei
Li, Hang
Li, Hengjie
Li, Yang
Liu, Yufei
Lu, Yanan
Lv, Qi
Ma, Haoxiang
Pang, Jiangmiao
Qiao, Yu
Qiu, Zherui
Shen, Yanqing
Shi, Xu
Tian, Yang
Wang, Bolun
Wang, Hanqing
Wang, Jiaheng
Wang, Tai
Wei, Xueyuan
Wu, Chao
Xie, Yiman
Xing, Boyang
Yang, Yuqiang
Yang, Yuyin
Yu, Qiaojun
Yuan, Feng
Zeng, Jia
Zhang, Jingjing
Zhang, Shenghan
Zhang, Shi
Zhaxi, Zhuoma
Zhou, Bowen
Zhou, Yuanzhen
Zhou, Yunsong
Zhu, Hongrui
Zhu, Yangkun
Zhu, Yuchen
Robotics
Prevalent Vision-Language-Action (VLA) models are typically built upon Multimodal Large Language Models (MLLMs) and demonstrate exceptional proficiency in semantic understanding, but they inherently lack the capability to deduce physical world dynamics. Consequently, recent approaches have shifted toward World Models, typically formulated via video prediction; however, these methods often suffer from a lack of semantic grounding and exhibit brittleness in the presence of video prediction errors. To synergize semantic understanding with dynamic predictive capabilities, we present InternVLA-A1. This model employs a unified Mixture-of-Transformers architecture, coordinating three experts for scene understanding, visual foresight generation, and action execution. These components interact seamlessly through a unified masked self attention mechanism. Building upon InternVL3 and Qwen3-VL, we instantiate InternVLA-A1 at 2B and 3B parameter scales. We pre-train these models on heterogeneous data sources over real-world robot data, synthetic simulation data, and human videos, covering over 692M frames. This hybrid training strategy effectively harnesses the diversity of synthetic simulation data while minimizing the sim-to-real gap. We evaluated InternVLA-A1 on 12 real-world robotic tasks and a simulation benchmark. The results show that InternVLA-A1 consistently outperforms prior leading models: compared with pi0.5, it achieves +4.4\% on static manipulation tasks and +2.6\% on the RoboTwin 2.0 simulation benchmark, and delivers a +26.7\% boost on dynamic manipulation tasks.
title InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation
topic Robotics
url https://arxiv.org/abs/2601.02456