Skip to main content
Back to timeline
arXivSource publication:

Fixed-gain PD forces a trade-off between motion and contact force: ImpedanceMimic lets the policy set joint impedance, tracking force better across 129 participants

Synopsis

Physics-based motion imitation usually converts joint-position targets into torques with a fixed-gain proportional–derivative (PD) controller, so one motion can correspond to different contact forces, and fixed gains struggle to serve stance and swing at once. ImpedanceMimic has a reinforcement-learning policy output per-joint stiffness, damping and position setpoints, supervised in training by force-plate ground reaction forces and center of pressure, while inference uses motion alone. Across 129 participants and six activities it beats fixed-gain PD, direct torque, muscle actuation and existing imitation controllers on joint-angle error and force fidelity, and transfers zero-shot to unseen humans and a simulated humanoid robot.

Source-provided article image: ImpedanceMimic: Learned Impedance as Action Space for Force-Accurate Motion Imitation
arXiv

Interpretation

The imitation controller no longer emits only joint-position offsets; it emits per-joint stiffness, damping and position setpoints, making contact force a quantity the policy adjusts directly. Earlier physics-based imitation mostly kept fixed-gain PD, which turns position targets into torques; different gains can reproduce the same motion yet yield different contact forces. Across 129 participants and six activities, the policy beats fixed-gain PD, direct torque, muscle actuation and existing imitation controllers such as PHC and SimPoE on joint-angle error and force fidelity, approaching the dedicated kinematics-to-force regressor GaitDynamics in force accuracy.

The policy learns to modulate impedance from motion and force measurements, yet needs no force-plate signal at inference. Gains are expressed as multiples of inertia-scaled priors, so one policy drives subject-specific scaled models without anthropometric input. Training uses 107 subject-specific models and evaluation is zero-shot on 22 held-out participants (49–93 kg, 1.50–1.82 m) and a simulated Unitree G1 humanoid; mean survival is 98.1% on held-out subjects and 82.1% on the G1, against about 56.9% and 44.0% for fixed-gain PD and direct torque on the robot.

Perspective

The demonstrated setting is joint-actuated human models and one simulated Unitree G1 in simulation, trained and evaluated on nine public datasets with synchronized force plates, with the policy at inference receiving simulator state and a force estimate computed from reference motion.

The authors note that the reference lookahead and force estimator currently draw on up to 400 ms of future motion, so the estimator is not yet causal, and that the simulation omits actuator bandwidth, stiction and torque limits that hardware would expose. Every method also performs worse on lateral cutting, flexion and jumping, which the authors partly attribute to the datasets lacking arms and to the low share of non-cyclic training clips.