BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation
CoRL 2026 Spotlight. A fully native 1-bit VLA with ternary weights, built on BitNet b1.58 2B4T, designed for edge devices instead of relying on post-hoc compression.CoRL 2026 Spotlight。原生 1-bit VLA，权重三值化，基于 BitNet b1.58 2B4T，面向边缘设备设计，而非依赖事后压缩。
BibTeX
@misc{wang2025bitvla,
title = {{BitVLA: 1-bit Vision-Language-Action Models for Robotics Manipulation}},
author = {Hongyu Wang and Chuyan Xiong and Ruiping Wang and Xilin Chen},
year = {2025},
eprint = {2506.07530},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2506.07530},
}