A practical “cookbook” for vision-language-action models: which backbones, perception pipelines, and action predictors actually work for robots.