ImpedanceMimic: Learned Impedance as Action Space for Force-Accurate Motion Imitation
Abstract
Recovering the forces behind a motion is ill-posed: different controllers can reproduce the same movement while generating different contact forces. Physics-based motion imitation resolves this ambiguity through simulation, but relies on fixed-gain proportional-derivative (PD) control. We show that fixed gains impose a trade-off between motion and force tracking, as contact and swing place conflicting demands on joint stiffness and damping. We introduce ImpedanceMimic, a reinforcement-learning policy that modulates per-joint stiffness, damping, and position setpoints from motion and force measurements, but requires only motion at inference. Across 129 participants and six activities, ImpedanceMimic outperforms matched fixed-gain PD under every integration scheme, direct-torque, muscle-actuated and existing imitation controllers in kinematic and force fidelity, approaching a dedicated motion-to-force regressor. By expressing gains as multiples of inertia-scaled priors, a single policy controls subject-specific biomechanical models without explicit anthropometric input and transfers zero-shot to unseen humans and a humanoid robot. The learned impedance forms a low-dimensional representation of motor state, shared across bodies, that generalizes to withheld activities: stiffness scales with load, and damping anticipates unloading.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
ImpedanceMimic introduces a learned impedance action space that resolves the fundamental trade-off between force authority and stability in physics-based motion imitation, achieving force-accurate tracking that transfers zero-shot across diverse human morphologies and to humanoid robots. The paper provides a rigorous theoretical foundation for why fixed-gain PD controllers fail in contact-rich tasks and demonstrates that learned, state-dependent impedance is necessary for high-fidelity force recovery, offering a significant advance in the field of physics-based character control and biomechanical simulation.
Methodology
The paper introduces ImpedanceMimic, a reinforcement learning policy that outputs per-joint stiffness, damping, and position setpoints rather than fixed gains or raw torques. The core theoretical contribution is the derivation of two bounds: a lower bound on stiffness required to generate reference contact forces (Bound 1) and an upper bound on damping required for discrete-time stability (Bound 2). The authors demonstrate that fixed-gain PD controllers are fundamentally limited by the conflict between these bounds during contact vs. swing phases, whereas learned impedance can satisfy both by modulating gains state-dependently. The method uses inertia-scaled priors to make the controller morphology-agnostic, allowing zero-shot transfer to unseen human bodies and a humanoid robot. The use of a kinematics-to-GRF estimator for training supervision, which is discarded at inference, is a clever architectural choice that enables force-accurate imitation without force sensors at test time.
Results & evaluation
The experimental evaluation is extensive and rigorous. The authors utilize a large-scale dataset of 129 participants across six activities (walking, running, standing, lateral, flexion, jumping), sourced from AddBiomechanics and other public datasets. They compare against fixed-PD, direct-torque, muscle-actuated, and existing imitation controllers (PHC, SimPoE). The results show ImpedanceMimic outperforming baselines in both kinematic and force fidelity. Crucially, they demonstrate zero-shot transfer to 22 unseen subjects and a Unitree G1 robot, as well as generalization to withheld activities. The ablation studies effectively isolate the contribution of force supervision and inertia scaling. The comparison to a dedicated motion-to-force regressor (GaitDynamics) provides a strong upper bound reference, showing the physics-based approach achieves comparable force accuracy while maintaining dynamic consistency.
Reproducibility
The paper is highly reproducible. Code, checkpoints, and data splits are released openly. The authors provide detailed descriptions of the simulation setup (MuJoCo, Warp), training hyperparameters (PPO, 2 billion steps, 8192 parallel worlds), and data preprocessing steps. The use of public datasets (AddBiomechanics) ensures data availability. The clear definition of metrics (RMSE, Pearson correlation, survival rate) and evaluation protocols (seen/unseen subjects, withheld activities) allows for fair comparison.
Limitations
The primary limitation is the reliance on a kinematics-to-GRF estimator for training, which may introduce biases if the estimator is inaccurate for certain activities or subjects. The method requires a physics simulation, which may not perfectly match real-world dynamics (e.g., actuator delays, friction). The zero-shot transfer to the robot is limited to a simulated environment; hardware validation is not provided. The model is a no-arm Rajagopal model, which may limit applicability to tasks requiring upper body coordination. The computational cost of training (2 billion steps) is significant, though the inference cost is low.
Broader impact
This work has significant implications for robotics, biomechanics, and animation. By providing a method to recover physically consistent forces from motion data, it enables more realistic character animation and more accurate biomechanical analysis. The zero-shot transfer capability suggests a path toward general-purpose controllers that can adapt to new bodies without retraining. The insights into the trade-off between force authority and stability in PD control are valuable for the broader robotics community. The open benchmark will likely become a standard for evaluating force-accurate motion imitation. ImpedanceMimic introduces a learned impedance action space that resolves the fundamental trade-off between force authority and stability in physics-based motion imitation, achieving force-accurate tracking that transfers zero-shot across diverse human morphologies and to humanoid robots. The paper provides a rigorous theoretical foundation for why fixed-gain PD controllers fail in contact-rich tasks and demonstrates that learned, state-dependent impedance is necessary for high-fidelity force recovery, offering a significant advance in the field of physics-based character control and biomechanical simulation.
Author of this paper? Send a correction.