Physical intelligence has advanced rapidly in research labs, yet reliable deployment in everyday environments remains elusive. One central obstacle is the extrinsic dynamics of the real world: objects move, sensing is partial, and other agents evolve over time. In this talk, I will present my research on three capabilities robots need to thrive in dynamic worlds: reactive, predictive, and proactive. First, I will introduce bidirectional decoding, a sequential action inference algorithm that accelerates reactions to moving objects without policy retraining. Next, I will describe a counterfactual-inspired approach to world modeling, which enables robust prediction of future dynamics in unseen safety-critical conditions. Finally, I will characterize the structural asymmetry between two common classes of dynamic models, deriving a principled verification mechanism for proactive exploration. I will close with an outlook on a broader shift in physical intelligence, from passive imitation on static datasets to active self-improvement in open dynamic worlds.
