A robot can have a powerful model and still struggle with a simple task. The missing piece is often experience. Real world robot data is expensive to collect. Every new task can mean more teleoperation, more physical trials, more engineering and more time. This is where @axisrobotics is taking an interesting approach. Instead of relying entirely on physical demonstrations, Axis connects digital twins, simulation and real robot data into one continuous learning loop. A real workspace can be reconstructed as a digital twin. That environment can then be diversified across objects, camera views, lighting and different scene conditions. Simulation creates more task aligned experience. Real demonstrations ground that experience in physical reality. Then failures become new targets for the next round of data collection. The Booster results make this idea much more tangible. With only 10 real demonstrations for an earbud case task, the robot reached the target in 0/20 trials. After adding 50 Axis simulation trajectories: 17/20 trials reached the target. Axis also filtered 42,046 simulation episodes from 95 repositories, covering more than 3.2M frames, and combined them with real Booster data to build a Booster specific foundation model. The changes were interesting: 17.27 cm → 5.78 cm Mean wrist to target distance 43.9% → 88.2% Motion from the instructed arm ~95% reduction Non target arm movement And with only 30 demonstrations per task: 6/16 → 14/16 Ordered task execution My takeaway is simple. The interesting part isn't just that Axis can generate more robot data. It's that the system is designed to make the data compound. A failure creates a new collection target. A new task improves the embodiment prior. That improved prior helps with the next task. Then the cycle repeats. Capture → Simulate → Train → Deploy → Learn → Repeat. For Physical AI, that kind of learning loop could be just as important as the model itself. That's the part of Axis I'm watching closely.