The real skill in sim to real might be knowing what NOT to randomize. One quiet line in the latest @axisrobotics Weekly caught me: domain randomization now keeps articulated objects fixed while varying the rest of the scene. That distinction matters. Synthetic diversity is useful until the variation starts changing the structure of the skill you’re trying to teach. A policy learning drawers or cabinets should experience different surrounding conditions without constantly losing the stable interaction it needs to understand. So the problem isnt: how much variation can we generate? Its: which variables should change, and which should remain invariant? That turns domain randomization from a volume trick into an engineering decision about what the robot should learn to ignore. More synthetic data isnt automatically better. Better controlled variation is. https://t.co/NQvnEdHquc For sim to real, which matters more: maximum diversity or choosing the right invariants?