Reproduce “SACQ: Structure-Aware Codebook Quantization for Ultra-Low-Bit Vision-Language-Action Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haesung Lee, Insu Han
SACQ compresses VLA policy weights by tailoring codebook fitting to projection roles and allocating precision according to decoded-action error, then refining the quantized policy. On simulated LIBERO, it achieves 93.85% average success at 1.5 language/action bits per weight, versus 94.10% for BF16, while reducing resident memory from 5.677 GB to 1.175 GB. The results do not establish performance on physical robots.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.