What does this equation mean?

The dominant scheme for the last stage is impedance control rather than pure position control, and the distinction matters for exactly the reason a robot has to touch things. A pure position controller tries to force a joint to a commanded angle regardless of what resists it, which is dangerous the instant the arm meets an object, a surface, or a person, because the controller will keep applying more torque to fight the very contact it should be responding to. An impedance controller instead makes the joint behave like a programmable spring and damper around the commanded target: $\tau = K_p (q_d - q) + K_d (\dot{q}_d - \dot{q}) + \tau_{ff}$. 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.