Accelerating Vision-Language-Action Model Integrated with Action Chunking via Parallel Decoding

Wenxuan Song¹*, Jiayi Chen¹*, Pengxiang Ding²˒³, Han Zhao²˒³, Wei Zhao²,Zhide Zhong¹, Zongyuan Ge⁴, Jun Ma¹, Haoang Li¹‡

¹ IRPN Lab, HKUST(GZ) ² MiLab, Westlake University ³ Zhejiang University ⁴ Monash University

We propose parallel decoding framework for VLA models integrated with action chunking.

RationalVLA: A Rational Vision-Language-Action Model with Dual System

Wenxuan Song, Jiayi Chen, Wenxue Li, Xu He, Han Zhao, Can Cui, Pengxiang Ding, Shiyan Su, Feilong Tang,Donglin Wang, Xuelian Cheng, Zongyuan Ge, Xinhu Zheng, Zhe Liu, Hesheng Wang, Haoang Li

• We introduce a new benchmark, the Rational Manipula-tion (RAMA) with defective instructions in 6 dimensions.
• We propose Rational Vision-Language-Action model (RationalVLA). It is a dual system for the robotic armto perceive the environment, handle unseen and defective instructions effectively, and demonstrate performant manipulation capabilities.

CEED-VLA: Consistency Vision-Language-Action Model with Early-Exit Decoding

Wenxuan Song¹*, Jiayi Chen¹*, Pengxiang Ding²˒³†, Yuxin Huang¹, Han Zhao²˒³, Donglin Wang², Haoang Li¹‡

¹ IRPN Lab, HKUST(GZ) ² MiLab, Westlake University ³ Zhejiang University

• Wepropose CEED-VLA, a universal acceleration method for significant inference speedup while maintaining manipulation performance.
• We conduct a consistency distillation process to unlock the model’s capabilities of fast inference.
• We identify the inefficient iteration as the bottleneck of Jacobi decoding’s speedup and propose the early-exit decoding to solve it.