Awesome VLA & WAM Quantization
A curated, categorized list of over 180 research papers on quantizing Vision-Language-Action and World-Action models so embodied robot policies fit within memory, latency and energy limits.
This list organizes quantization research for robot policy models along a five-category taxonomy and a rough timeline, moving from numerical baselines originally developed for large language models, through VLA-specific post-training and quantization-aware training methods, to deployment and action- or world-model-specific quantization work. Each entry links to its paper on arXiv or a scholarly search when no arXiv listing exists, and landmark papers are marked separately from the rest.
A companion open-source toolkit, quant-vla, is planned for release alongside the list, combining post-training quantization with the LIBERO robot-manipulation benchmark for evaluating several existing quantized VLA approaches. The repository also tracks related robot policy models, world models, benchmarking environments, and a leaderboard scaffold for comparing results across the surveyed methods.
It is licensed under the MIT license.