Given the abundance of empirical laws in astrophysics, the rise of agnostic and automatic methods to derive them from data is of great interest. This concept is embodied in symbolic regression, which seeks to identify the best functional form fitting a dataset. Here we present a protocol for deducing both physical laws but also the constants of nature appearing in those with their associated units. Our method is grounded in the Physical Symbolic Optimization framework, which integrates dimensional analysis with deep reinforcement learning. We showcase our approach on a panel of equations from (astro)-physics.