Ω-QVLA: Robust Quantization for Vision-Language-Action Models via Composite Rotation and Per-step Scaling

Xinyu Wang, Mingze Li, Sicheng Lyu, Dongxiu Liu, Kaicheng Yang et al.|May 27, 2026arXiv

Key Takeaway

You can now quantize entire robot control models to 4-bit precision without performance loss, making them deployable on edge devices—previous methods thought this was impossible for the action generation part.

Summary

This paper introduces Omega-QVLA, a quantization method that compresses Vision-Language-Action models (which control robots) to 4-bit precision for both weights and activations.

efficiency multimodal

Key Terms

quantization vision-language-action-model diffusion-transformer post-training-quantization mixed-precision-training