One thing I find interesting about @axisrobotics is the focus on the data layer of robotics. Training a robot isn’t just about having a capable model. It needs diverse trajectories, different environments and enough interaction data to actually generalize. Using simulation to generate and augment that data could make the whole process much more scalable. The sim-to-real part is obviously where things get really interesting. Physical AI is still early, but the infrastructure being built around it is becoming just as important as the models themselves.