And nowhere does anyone explain exactly what “poisoned data” is and just how the claim that training a model on a small amount of such data will have a big impact.
With cheap generic robots coming this can be a real problem. Human supervision can help when there is one.