What does this equation mean?

Here $q_d$ and q are the desired and measured joint positions, $K_p$ and $K_d$ set how stiff and how damped that virtual spring-and-damper is, and $\tau_{ff}$ is a feedforward term — gravity compensation, or a reaction-force term handed down from a higher-level controller — added on top. The learned policy’s contribution to this equation is $q_d$ : a target, updated at whatever rate the policy runs. Everything else in the equation runs at a much higher rate, underneath the policy, and it is what actually decides how the robot responds to unexpected resistance in the interval between two policy decisions. Lower the gains and the joint yields to contact instead of fighting it; that yielding, not any property of the learned policy, is most of what keeps a light unexpected collision from becoming a hard one.