opengalaxea/galaxeavla

GalaxeaVLA is a pretrained vision-language-action model for robot control. Its main runtime path turns robot observations and task instructions into batched inference, then postprocessed action chunks for execution; the README describes closed-loop replanning from new observations. The repository also supports demonstration-data loading and processing, action-token representation, model training/fine-tuning, and simulation evaluation. The tree and excerpts establish these subsystems, but do not expose every training or serving call site, so uncertain internal wiring is omitted.