The deeper I reflect on robotics data, the more convinced I become that simply accumulating more data isn't necessarily the solution. The real question is whether we are collecting the right data This is what caught my attention about the direction behind @axisrobotics. A robot model continues to advance and get better, while the information surrounding it may remain unchanged. That creates a problem Some behaviors may already be easy for the model, so collecting more of them adds very little At the same time, those rare cases where the model continues to struggle might actually matter way more. Axis is reimagining the workflow by shifting away from the traditional model of collecting data first and training afterward, instead pursuing a different cycle: Collect Train Evaluate Adapt the data Train again The model becomes part of the data engine Its failures can show us what is missing Its performance can help identify which existing data is still useful. And human corrections can provide new training signals exactly where the model struggles I'm genuinely excited about this concept because it shifts the focus away from just expanding dataset size and toward creating a more intelligent approach to how data is organized and distributed. More data does not automatically mean better learning Every so often, the most precious information comes from that tiny fraction of data that helps the model learn something it hasn't quite mastered yet. That is where Physical AI gets really interesting to me The model is not just learning from the data anymore It enables us to predict which information will follow next.