ImpedanceMimic: Learned Impedance as Action Space for Force-Accurate Motion Imitation
Matthias Jammot, Pembe Gizem Özdil
cs.RO, cs.GR, cs.LG
2026-11-14
Learned per-joint stiffness and damping reach 2.28° joint RMSE and 0.80 vertical GRF correlation on 129 people, using only motion at inference.
The same joint motion can come from many torque trajectories, so foot force is not fixed by kinematics. Force plates stay in the lab. Insoles drift and miss shear. A network that regresses ground reaction force from motion can still load the floor during flight and put the center of pressure outside the foot. That map also cannot drive a simulated body.
Tracking a recorded motion in a physics engine makes contact forces agree with the contact model. Standard imitation policies emit joint position targets, and a fixed-gain PD controller turns position error into torque. A 2017 action-space comparison found fixed-gain PD best at following the pose, and later character controllers kept it. Scoring force removes that default. Contact wants a stiff joint, because stiffness sets how much force the PD can produce. Swing wants light damping, because the joint only moves the foot. One fixed pair of gains has to favor one phase.
Near the reference pose, the position offset is the policy's only torque lever. Stiffness must cover the inverse-dynamics torque divided by that offset, or a fixed-gain controller drops either the pose or the force. At a 2 ms physics step, damping must stay under about 2I/h or the discrete loop is unstable. Swing inertia is only the foot, so a damping gain that is fine in contact can slam the torque into its limits after toe-off. On a single hinged rod the two bounds are 32 Nm/rad and 20 Nm·s/rad.
The action is a stiffness residual, a damping residual, and a position offset on each of 15 actuated joints, growing the action from 15 to 45 numbers. Gains are exponential multiples of a rest-pose inertia prior, with log-scales 1.25 and 1.5, so they stay positive. The same multiple gives a similar closed-loop frequency on another body. The policy never sees height, mass, or a subject id. A 50 ms lag smooths gain changes, which otherwise show up as force spikes, and it forces the joint to stiffen before touchdown.
Test time uses no force plate. The actor sees 215 numbers: simulator proprioception, plus joint error, estimated force, and foot pose at 10, 50, and 250 ms ahead. The estimator splits whole-body force from center-of-mass acceleration, then a frozen residual network corrects it. Vertical correlation is 0.979 on seen people and 0.943 on unseen people. Measured force is used only in training. The actor never receives it at inference. The force-reward weight rises from 0.25 to 0.5.
The pool is 9 public datasets, 129 people, and 20.2 hours, from 44 to 126 kg and 1.46 to 1.95 m, across walking, running, standing, lateral movement, flexion, and jumping. Each person is a scaled armless Rajagopal model in MuJoCo Warp. PPO runs for 2 billion steps on 8,192 worlds. One run takes about 1.3 GPU-hours.
On held-out clips from people seen in training, the six-activity means are below. GaitDynamics only regresses force, so it is not ranked as a controller.
| Method | Survival | Joint RMSE | Vertical GRF r | Lateral GRF r |
| GaitDynamics | n/a | n/a | 0.91 | 0.92 |
| Fixed-gain PD | 93.0% | 2.91° | 0.73 | 0.53 |
| Direct torque | 95.4% | 4.48° | 0.82 | 0.62 |
| Muscle excitation | 84.8% | 7.18° | 0.77 | 0.62 |
| ImpedanceMimic | 96.9% | 2.28° | 0.80 | 0.70 |
Walking and running reach vertical correlations of 0.98 and 0.99, with joint errors of 1.66° and 2.24°. Standing is 0.56, because double support does not decide how the feet split the load. Direct torque is slightly higher on vertical correlation, 0.82 against 0.80, mostly in standing, at about twice the joint error. Emitting torque on every physics step leaves that error at 4.55° against 4.48°. Removing the force reward lowers joint error and vertical correlation in every action space.
On 22 unseen people (49 to 93 kg), survival is 98.1%, joint error 2.45°, and vertical correlation 0.76. A simulated Unitree G1 at 33 kg and 1.32 m, lighter and shorter than all 107 training subjects, drops from 98.1% to 82.1% survival. Fixed PD and torque fall to 56.9% and 44.0%. Held out by activity, survival is 89.6% and vertical correlation 0.72, about 7 survival points below training on all six activities, against 16 to 23 for the baselines. Withholding walking drops vertical correlation from 0.98 to 0.66.
Ankle stiffness moves from 0.6 times nominal at heel strike to 1.5 times at push-off. Its contact average tracks peak vertical force at r = 0.93. Damping peaks 128 to 196 ms before toe-off and is about one third of nominal at toe-off. A classifier on the first ten principal components reads the activity from one cycle of gains at 92.5% macro recall. Under implicit damping, where the damping term sits inside the integrator, joint error is 2.22° against 8.66° for the existing imitator PHC.
The regressor still loads the ground on 30% of lateral-cutting flight frames and 39% of jump flight frames. ImpedanceMimic does so on 11% and 17%, a contact-timing error. The center of pressure leaves the foot on 6% and 17% of loaded frames for the regressor, and on at most 0.2% here.
Fixed-gain PD became the default action space because pose tracking was the score that mattered. On a shared force benchmark, those gains buy the pose by giving up the force. Learned impedance keeps the PD feedback at every physics step.
One policy covers 129 people, six activities, and per-foot contact force, and it reads only reference motion at test time. Transfer across bodies comes from the inertia prior. One cycle of gains identifies the activity at 92.5% macro recall, with no kinematics and no force in that decoder. Powered prostheses already schedule joint impedance by gait phase, with gains tuned by hand for each user. A state-dependent schedule could replace that tuning only after the force estimate is causal. The current lookahead uses up to 400 ms of future motion.
Vertical correlation still trails a dedicated regressor. What comes back is a contact force that joints and the foot model can actually produce, on a body the controller was not trained on.
The model has no arms, so arm swing is not available for balance. Non-cyclic clips are 16% of training, jump survival is 83.8%, and standing vertical correlation is 0.56, against 0.75 for the regressor.
The simulator omits actuator bandwidth, stiction, and torque limits. A shorter hardware step raises the damping ceiling, so the learned damping is conservative. The G1 numbers are retargeted simulation, and the reference force is only the center-of-mass prior.
Ankle stiffness rises less steeply than the 2014 human measurements. Quiet standing supplies 0.4 of the stiffness needed to resist toppling, against about 0.9 in people. Gains only modulate around the chosen prior. Under stable PD, which evaluates feedback at the next state, survival is 93.8%, below 96.6% for fixed PD at doubled priors. The gap is in flexion, 85.1% against 99.0%. Doubling the prior never raised vertical correlation. The policy also reads reference motion 10, 50, and 250 ms ahead. Online control without that future was not tested. Removing SimPoE's pelvis wrench drops survival to 85%, 80%, and 56%. That wrench supplied more than half of the mediolateral force.