Physical Intelligence published research showing its π0.7 model can direct robots to perform tasks they were never explicitly trained on by combining skills learned in different contexts — a capability called compositional generalization. The model successfully used unfamiliar appliances like air fryers and performed complex tasks such as making coffee and folding laundry, even surprising the researchers who didn't expect this level of knowledge transfer from the training data. The researchers acknowledge the model still requires step-by-step verbal coaching for new tasks and can't yet execute multi-step sequences from a single high-level command, and they've measured performance against their own specialist models rather than independent benchmarks. This suggests robotics AI may be approaching a scaling inflection point similar to large language models, where capabilities compound faster than underlying training data would predict.