ImpedanceMimic: Learned Impedance as an Action Space for Force-Accurate Motion Imitation

Recovering the forces behind human motion from its kinematics is ill-posed: the same movement can produce different ground reaction forces (GRFs). Physics-based imitation resolves this through a contact model and a torque trajectory, yet existing pipelines optimize kinematic fidelity alone, fixing joint impedance and tuning only the position setpoint. We introduce ImpedanceMimic, a state-dependent policy that recovers per-foot GRFs through learned impedance, anchored to inertia-scaled priors and trained against force-plate GRF and centre-of-pressure measurements. Against fixed-PD and direct-torque actuation, it is the only action space that achieves both kinematic and force accuracy ($q_{\text{rmse}}$ = 1.70°, $F_{Z_r}$ = 0.96) with the smoothest torques (7.6 Nm/kg/s jerk). Trained subject-blind on 11 morphologies, a single policy walks a held-out subject zero-shot (episode completion 0.96 vs. 0.61 for fixed-PD), acting as a shared controller across embodiments.