Reshaping Action Error Distributions for Reliable Vision-Language-Action Models

Shuanghao Bai, Dakai Wang, Cheng Chi, Wanqi Zhou, Jing Lyu, Xiaoguang Zhao, Pengwei Wang, Zhongyuan Wang, Lei Xing, Shanghang Zhang, Badong Chen, [0.4em] Institute of Artificial Intelligence, Robotics, Xi’an Jiaotong University · 2026

Read Reshaping Action Error Distributions for Reliable Vision-Language-Action Models by Shuanghao Bai, Dakai Wang, Cheng Chi, Wanqi Zhou, Jing Lyu, Xiaoguang Zhao, Pengwei Wang, Zhongyuan Wang, Lei Xing, Shanghang Zhang, Badong Chen, [0.4em] Institute of Artificial Intelligence, Robotics, Xi’an Jiaotong University (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-2602-04228.txt.

Description

Shaping the whole error distribution beats punishing each mistake alone. Shuanghao Bai and colleagues apply that idea to vision-language-action robot models in a 2026 paper. Standard training uses pointwise losses, mean squared error for continuous actions, which force every prediction toward its target individually.

Their approach adds minimum error entropy, an information-theoretic objective, at trajectory level, with two weighted variants alongside the usual loss. Tests spanned standard, few-shot, and noisy settings on leading architectures, in simulation benchmarks and real manipulation tasks. Reshaping errors should make policies degrade gracefully when the world stops matching training.

Catalog record

CREATORShuanghao Bai, Dakai Wang, Cheng Chi, Wanqi Zhou, Jing Lyu, Xiaoguang Zhao, Pengwei Wang, Zhongyuan Wang, Lei Xing, Shanghang Zhang, Badong Chen, [0.4em] Institute of Artificial Intelligence, Robotics, Xi’an Jiaotong University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-04228

Browse the full library · About the Tower of Babel