Motion-only imitation learning can track human gait kinematics well and still get ground reaction forces, center of pressure and joint moments wrong. Adding GRF and CoP to the reward fixes much of it.