Source item

@iyanmoonyang @ryancjulian even worse in robotics than llm. eval depends on hardware, objects, environment (scene, lighting, calibration), policies / recovery behavior… pairwise preference feels tough to do