acceptodds
Under review as a conference paper at ICLR 2027

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 exerting different contact forces. Most motion-imitation methods anchor the humanoid in a physics simulator that grounds the contact dynamics, and drive it with a fixed-gain proportional–derivative (PD) controller. We show that fixed gains enforce a compromise between tracking the motion and recovering the resulting forces, where contact and free swing place opposing demands on the gains. To overcome this limitation, we present ImpedanceMimic, a motion-imitation reinforcement-learning policy that outputs per-joint stiffness, damping and position setpoints, supervised by force measurements withheld at inference. Across 129 participants, nine datasets and six diverse activities from lunges to jumping, ImpedanceMimic outperforms fixed-gain PD, direct torque and PHC on kinematic and force fidelity alike, and approaches a dedicated motion-to-force regressor. With gains learned as multiples of inertia-scaled priors, our policy is the first to drive subject-scaled biomechanical models for every participant, without anthropometric input, and transfers zero-shot to new bodies. The learned impedance forms an activity-specific, morphology-invariant signature of the underlying motor strategy, allowing the policy to generalize to unseen activities. This signature occupies a low-dimensional, interpretable space where stiffness scales with contact force and damping anticipates contact, consistent with human perturbation studies. We release our code, model checkpoints and benchmarks at https://impedancemimic.github.io/ImpedanceMimic.

Primary area:
applications to robotics, autonomy, planning
Keywords:
motion imitation, impedance control, reinforcement learning, action spaces, contact forces, biomechanical simulation, humanoid control, physics simulators

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_H7g4OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper argues that with explicit PD control, fixed gains cannot meet both the force authority needed in contact (Bound 1) and the stability needed in swing (Bound 2), so the controller gives up either pose or force. It proposes ImpedanceMimic, a policy that outputs per-joint impedance and is trained with force-plate GRF and CoP rewards. On 129 participants and six activities it reports better motion and force tracking than fixed-gain PD, better motion tracking and survival than direct torque, and transfer to unseen bodies and activities. The learned gains are interpreted as an activity-specific physiological signature.

Critical strengths

The paper scales imitation with measured GRF and CoP rewards from a single walking participant to 129 participants and six activities, without a residual wrench. It also evaluates forces in detail, which most physics-based imitation work does not, and its physical-consistency metrics show where regression models produce impossible forces. I think this evaluation setup will be useful for future work.

Critical weaknesses

W1. Much of the claimed contribution seems to come from PD control itself rather than from the learned impedance, and stable or implicit PD already avoids the problems the paper sets up.

The two bounds are derived for PD applied as an explicit force (Section 3.1). Stable PD has no Bound 2, and because its torque accounts for gravity and external forces through the predicted acceleration, Bound 1 does not hold in the same form either (Tan et al., 2011). Widely used imitation controllers avoid explicit PD: DeepMimic uses stable PD and PHC uses Isaac Gym's implicit PD. It is unclear whether the analysis applies beyond explicit PD.

The generalisation results raise the same question. Unseen subjects come from the same corpus, and fixed-gain PD with the same inertia prior already survives 91.3% of their clips (Table 1), so most of the transfer seems to come from PD control plus the prior. For unseen activities, the paper attributes the failure of fixed-gain PD to instability, again a property of explicit PD, whereas PHC with implicit PD reports high success on motions outside its training set.

Because PD control and the learned gains are not separated, it is hard to tell what the method itself contributes.

W2. The comparisons are not like for like, so the advantage over the baselines is not established.

Direct torque has a higher FZr than ImpedanceMimic in Table 1, although the abstract claims the opposite. The two use different pose kernels, and removing amplitude normalisation, which is what distinguishes Torque's kernel, raises ImpedanceMimic's FZr from .80 to .85, above Torque's .83 (Table 19). Torque also has lower hip and knee torque limits and holds its torque for 10 ms, so Table 1 cannot separate the action space from these choices. Joint torques are also unresolved: ImpedanceMimic's RMSE against inverse dynamics is higher than Torque's (Table 10), and the metric is pooled over nine joints, so it does not show how well the knee, the joint behind the paper's motivation, is reproduced.

The PHC baseline appears to under-perform. PHC reports 98.9% success and 26.9 mm MPJPE on 11,313 AMASS sequences (Luo et al., 2023), but the port here has an 11.8° joint-angle error, four times that of fixed-gain PD on the same bodies. A controller that optimises only kinematic imitation should not track kinematics five times worse than ImpedanceMimic, which suggests the port does not reproduce PHC.

SimPoE and ReGAIL, the close methods that learn PD gains, are discussed but not compared. ImpedanceMimic's gains (Eq. 3) are also per-joint multipliers on base gains, so a direct comparison is needed to show what the inertia-based gains and the force reward add.

W3. The evidence does not support reading the learned gains as human impedance.

Human joint impedance is modulated largely by co-activation of antagonist muscles (Hogan, 1984). The same net torque can come with very different impedance, so kinematics and GRF alone cannot identify it. The values also differ from the cited measurements: the stance-phase stiffness rise is about one seventh of that in Rouse et al. (2014), and quiet-standing stiffness is 0.4 of the toppling stiffness, against about 0.9 in Loram & Lakie (2002). Calling them human is too strong for a stated contribution.

Tan, J., Liu, C. K., Turk, G. Stable Proportional-Derivative Controllers. IEEE CG&A 31(4):34–44, 2011.

Hogan, N. Adaptive control of mechanical impedance by coactivation of antagonist muscles. IEEE TAC 29(8):681–690, 1984.

Questions for authors

Q1 (W1). How does a fixed-gain baseline with stable PD or implicit damping, with high stiffness and adequate damping, compare on the seen, unseen and leave-one-activity-out splits? A clear lead for ImpedanceMimic there would make the contribution of the learned gains clear.

Q2 (W2). Could you compare against Torque with the same pose kernel and torque limits as the PD variants, and at a higher policy rate?

Q3 (W2). Could you report joint-torque RMSE and correlation per joint, at least for the knee and ankle?

Q4 (W2). Does the PHC port reproduce the published results in its original setting (AMASS)? If not, the PHC rows cannot support the comparison.

Q5 (W2). Could you compare with SimPoE's meta-PD and ReGAIL's gain and damping actions on the same data and reward?

Q6 (W3). The gain analysis uses only the seed-0 policy. Does the signature hold across seeds and gain ranges?

7 comments
AuthorsOpenReview ↗

[1/5] W1, Q1. Fixed and learned gains under stable PD and implicit damping

Dear reviewer,

Thank you for this detailed review. Your questions led to the largest additions to the revision, and every experiment you asked for is now in the manuscript or in this thread, with 5 seeds unless stated. Table and section numbers refer to the revised manuscript.

We ran the experiment you proposed in Q1. When both methods use the same integration scheme and the same gains, ImpedanceMimic tracks joint angles and all three force components better than fixed gains in every comparison, and survives longer in all but one. Stiffer fixed gains improve joint angles, but never the vertical force correlation.

Do the bounds apply beyond explicit PD? (W1)

You are right that Bound 2, the damping limit, applies only to explicit PD, and §3.1 now attributes it to explicit PD. Bound 1 still constrains fixed gains under the other two schemes. Implicit damping leaves the spring term unchanged and stable PD only evaluates it one step ahead, so a fixed stiffness still limits the available torque (App. A.2.3). We kept explicit PD as the main setting because it puts all four action spaces on the same integrator and gives the bounds in closed form, and we added the other two schemes as you suggested.

Fixed gains under stable PD and implicit damping (Q1)

We retrained Fixed-PD and ImpedanceMimic under stable PD and implicit damping, each with the nominal gains and with doubled gains, which provide the "high stiffness and adequate damping" you asked for (Table 2, App. A.2.3). Under explicit PD, only the stiffness is doubled, since doubled damping would violate Bound 2. Comparing the two methods at the same scheme and gain level gives 6 pairs (3 schemes × 2 gain levels) on 3 benchmarks, or 18 comparisons.

Same scheme and gainsImpedanceMimic better in
Joint-angle error18 of 18
Vertical force correlation18 of 18
Lateral force correlation18 of 18
Anterior–posterior force correlation18 of 18
Survival17 of 18

The one exception is survival on seen subjects at doubled gains under stable PD (96.3 against 96.6 %, within one seed sd), which is also where the margins are smallest. The lead in vertical force correlation is .06 to .13 under explicit PD, .04 to .05 under implicit damping and .01 to .04 under stable PD (Table 2). In survival, the lead under stable PD and implicit damping is small on seen subjects (at most 1.5 points) and grows away from the training data, to 2.2 to 11.8 points on withheld activities (Table 2) and 1.4 to 14.9 points on the G1 robot (Table 14).

Two further results show whether stiffer fixed gains could close the gap.

  • Doubling the gains never raises the vertical force correlation. It lowers the joint-angle error of fixed gains, but the vertical force correlation stays the same or drops, for both methods and under every scheme. Stiffer fixed gains therefore improve motion tracking without improving vertical force tracking.
  • The choice of gain scale still matters. Under stable PD with the nominal gains, ImpedanceMimic survives less often than Fixed-PD with doubled gains, mainly in flexion (85.1 against 99.0 % on seen subjects). With doubled gains, ImpedanceMimic closes this gap (98.4 %). The learned gains adjust around the chosen gains rather than replace that choice, as §4.4 now states.

Why fixed gains fall short. Stable PD and implicit damping remove the damping limit, but not the need to change stiffness with load. Both the torque a joint can produce in contact (Bound 1) and how the load is split between the feet in double support depend on the stiffness, which fixed gains cannot adapt (§4.4). The learned gains raise the stiffness with load (§4.3), which fixed gains cannot do under any scheme.

The picture is the same under a 20 ms observation delay, without retraining. Matched fixed gains lose 1.6 to 2.1 times as much survival as the learned gains under stable PD and implicit damping, and 3.5 to 8 times as much under explicit PD (Table 24).

AuthorsOpenReview ↗

[2/5] W1. Transfer to unseen subjects and withheld activities

We agree that survival on unseen subjects comes largely from PD control, and that the integrator explains part of the failure on withheld activities. What the learned gains add is force accuracy on unseen subjects, and the smallest loss of survival on withheld activities.

Does transfer to unseen subjects come from PD control and the prior? (W1)

Largely from PD control, yes. The unseen subjects come from the same corpus, and with stable PD or implicit damping, doubled fixed gains reach 96.2 and 96.5 % survival on them (Table 2). The prior matters little for survival. Without it, ImpedanceMimic still reaches 97.5 % survival, but its torque jerk rises 2.7-fold. Activity means on the 22 unseen subjects over 5 seeds:

Unseen subjectsSurvival (%)Joint-angle RMSE (°)Vertical GRF rTorque jerk
Fixed-PD with the prior (Tables 1, 21)91.53.13.7056.4
ImpedanceMimic without the prior (Table 27)97.52.68.7553.7
ImpedanceMimic98.12.45.7619.9

What the learned gains add is force accuracy. Compared with fixed gains at the same scheme and gain level, they raise the vertical force correlation on unseen subjects by .02 to .07 and the lateral one by .02 to .20 (Table 2).

Beyond the corpus, the same policies control a simulated Unitree G1 humanoid zero-shot, a body lighter and shorter (33 kg, 1.32 m) than every training participant. ImpedanceMimic reaches 82.1 % survival, against 56.9 % for Fixed-PD and 44.0 % for Torque (Table 13), and survives longer than fixed gains at the same scheme and gains in every case. Under stable PD, fixed gains reach a slightly higher vertical force correlation on the robot (.45 against .43, Table 14).

Is the failure on withheld activities a matter of stability? (W1)

In part, and the revision no longer attributes it to instability. With stable PD and implicit damping, Fixed-PD reaches 78.9 and 80.8 % survival on withheld activities instead of 69.8 % under explicit PD, and 85.8 and 83.4 % with doubled gains (Table 2).

The integrator does not explain everything, though. Every controller still loses survival when an activity is withheld from training.

Survival, seen → withheld (%)ExplicitStableImplicit
Fixed-PD, 2×96.2 → 83.196.6 → 85.895.8 → 83.4
ImpedanceMimic96.9 → 89.693.8 → 83.597.2 → 92.6
ImpedanceMimic, 2×96.9 → 90.396.3 → 88.096.8 → 90.4
PHC93.5 → 67.2
PHC, our reward93.5 → 70.2

Among controllers that act only through their joints, ImpedanceMimic loses the least under every scheme and at either gain level. The Discussion names two further factors, the small share of non-cyclic activities in training (16 % of clips) and subject models without arms.

PHC's high success outside its training set. This success refers to AMASS-Test, which holds out sequences from the Transitions and SSM subsets, not a whole activity. Our implementation imitates 92.2 % of these 140 motions (Table 10), but reaches only 67.2 % survival on our withheld-activity benchmark with its own reward, and 70.2 % with ours (Tables 2, 21).

AuthorsOpenReview ↗

[3/5] W2, Q2, Q3. Torque variants and per-joint torques

With the same pose kernel ImpedanceMimic beats Torque on every force metric, and neither the PD torque limits nor a 500 Hz policy rate improve Torque.

Torque's vertical force correlation and the abstract (W2)

You are right that Torque has the higher vertical force correlation on seen subjects (.82 against .80). §4.1 now says so, and notes that the lead comes mostly from standing, where the force stays close to body weight. On the other force metrics, which "force fidelity" in the abstract also covers, ImpedanceMimic is more accurate. Its lateral force correlation is higher on every benchmark (.70 / .65 / .57 against .62 / .53 / .47), as are its anterior–posterior force correlation and its center-of-pressure accuracy. On vertical force the two are otherwise close, with equal RMSE and correlations of .76 against .79 on unseen subjects and .72 against .71 on withheld activities. ImpedanceMimic also survives more often with half the joint-angle error (2.28 against 4.48°).

Torque with the same pose kernel, torque limits and a higher policy rate (Q2)

We ran all three (Table 22, seen subjects).

Survival (%)Joint-angle RMSE (°)Vertical GRF rLateral GRF rAP GRF r
Torque95.44.48.82.62.54
ImpedanceMimic with Torque's kernel97.62.86.85.75.65
Torque with ImpedanceMimic's kernel82.844.9.60.49.43
Torque with the PD torque limits82.553.4.09.04.06
Torque at 500 Hz91.94.55.65.50.41
  • Same pose kernel. With Torque's kernel, ImpedanceMimic beats Torque in survival, joint angles and every force metric. Torque cannot be trained with ImpedanceMimic's kernel. ImpedanceMimic keeps its own kernel because it gives a lower joint-angle error (2.28 against 2.86°) at a small cost in force accuracy, and we report both versions (Tables 22, 28).
  • Same torque limits. Torque's own output scale already lies below the PD torque limits at the hip, knee and ankle for nearly all subjects (§3.2). With the PD limits, Torque fails to learn to track.
  • Higher policy rate. Acting at every 2 ms physics step over the same simulated time, Torque reaches 4.55° against 4.48° at 100 Hz, so holding the torque for 10 ms is not what limits it. A torque policy has to learn the feedback that the PD law computes at every step (Peng and van de Panne, 2017).

We report the two failed variants as the outcome of the settings you requested, not as tuned baselines.

Per-joint torques (Q3)

Inverse-dynamics torques are themselves model-based estimates, and the goal of the paper is force-accurate motion imitation, so we report them for transparency and as a reference for future work. Seen subjects, 5 seeds, all six activities / without standing, RMSE in Nm/kg.

ImpedanceMimicTorqueFixed-PDMuscleImpedanceMimic, Torque's kernel
Knee RMSE.65 / .46.48 / .47.75 / .70.64 / .47.55 / .44
Knee r.62 / .74.61 / .73.50 / .60.57 / .69.63 / .76
Ankle RMSE.40 / .45.35 / .37.64 / .72.37 / .40.39 / .43
Ankle r.57 / .63.61 / .69.44 / .48.52 / .60.59 / .65
9 joints, RMSE.62 / .57.54 / .58.74 / .78.59 / .55.56 / .55

Torque's lower knee error comes from standing (0.55 against 1.60 Nm/kg for ImpedanceMimic), where the reference knee rests at its 0° stop. Without standing, the pooled RMSE is equal (.57 against .58), and ImpedanceMimic matches or exceeds Torque in knee correlation. At the ankle, Torque is more accurate, mainly in jumping (.43 against .75). Unseen subjects show the same pattern.

Reviewer_RKQpOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ImpedanceMimic to address the task of physics-based human motion imitation using reinforcement learning and an adaptive PD controller for computing the joint-space torques. The reconstructed physic-based motions are expected to produce the same ground reaction force as measurements from force plate sensors, if the imitation is successful. The experiments are conducted on the AddBiomechanics dataset, consisting of motion captures and synchronized ground reaction forces and center of pressure data.

Critical strengths
  • Compared to fixed-PD setups from previous methods, the proposal of trainable/adaptive PD shows clear advantages, especially towards the ability to complete a simulation on more challenging actions such as jumping.
  • The simulated ground reaction forces match well to reference measurements from force plate sensors.
Critical weaknesses
  • The base values of PD, k0pk_{0}^{p} and k0dk_{0}^{d}, are computed directly from the respective inertia IjjI_{jj}, which creates a strong dependency on GT mass measurement of the people. This information is not easy to obtain especially from in-the-wild data, which heavily limits the applicability of ImpedanceMimic downstream.
  • In Sec. 3.2, it is unclear why there is a need for the input of anticipated contact force F^t\hat{F}_t, and later on it becomes F^aux\hat{F}^{aux} in the objective function in Eq. 8. How is this anticipated contact force comparable to the resulting FF from the simulation and the reference GT force plate data? And if F^\hat{F} is already good enough, would there be a need for the whole simulation pipeline of motion imitation for obtaining simulated contact forces?
  • In Eq. 1, there exist the selection matrix SS that remains unexplained. According to MuJoCo documentation, the joints are individually controlled and evaluated. Please clarify if this selection matrix is identity matrix, or an another output from ImpedanceMimic.
Questions for authors

Q1: Why the simulation step size is chosen as 2ms2ms? Is this a requirement for stability, and would it be better to match it with the motion capture step of 10ms10ms?

Q2: What is the specific value of the scaling factor α\alpha in Sec. 3.1?

Q3: What is the difference between the contact force λ\lambda in Eq. 1 with the simulated FF?

3 comments
AuthorsOpenReview ↗

[1/3] Dependence on body mass and the inertia-scaled prior

Dear reviewer,

Thank you for these precise questions, which helped us clarify the method, and for recognizing the advantage of learned impedance in demanding activities such as jumping. This advantage does not depend on the integration scheme. We retrained Fixed-PD and ImpedanceMimic under stable PD and implicit damping, and with the same nominal gains, ImpedanceMimic improves on Fixed-PD in survival, joint-angle error and force correlation under every scheme and on every benchmark (Table 2). We also clarified the presentation, separating the four force quantities of the method (§3.2), defining the selection matrix and the contact dimensions (App. A.2.2) and retitling §4.3 to a narrower claim. Table and section numbers refer to the revised manuscript, and author–year citations to its reference list, except for the two new references listed below.

The prior comes from the body model that any physics-based imitation needs, it is not required for transfer, and subject-specific models can already be obtained from video.

Does the dependence on body mass limit in-the-wild use? (W1)

We agree that the prior needs the inertia of each individual, and offer three clarifications.

Every physics-based imitation method needs a body model. Simulating an individual requires segment masses and inertias. Our prior comes from this same model, and the policy itself receives no anthropometric input.

The prior is not required for transfer. Retrained without inertia scaling, with one gain set for all bodies, ImpedanceMimic transfers to the 22 unseen subjects with nearly unchanged survival and force accuracy (Table 27, activity means over 5 seeds).

Unseen subjectsSurvival (%)Joint-angle RMSE (°)Vertical GRF rTorque jerk
ImpedanceMimic98.12.45.7619.9
Without inertia scaling97.52.68.7553.7

Without the prior, the torques become less smooth (2.7 times the torque jerk), but transfer is unaffected. The same policies also control a simulated Unitree G1 humanoid zero-shot, with a prior derived from the robot's own model (Table 13).

Subject-specific models can already be obtained from video. The procedure is the same as in the laboratory, where a static calibration pose and the measured body mass scale a generic model. OpenCap Monocular (Gilon et al., 2026) estimates a scaled OpenSim musculoskeletal model and its kinematics from a single smartphone video, and SMPL2AddBiomechanics, released with SKEL (Keller et al., 2023), fits a scaled OpenSim skeleton to the SMPL motion of video-based pose estimators. Body mass, unlike ground reaction force, is easy to measure outside the laboratory. These tools output a scaled model and its reference kinematics, which could serve as inputs to our model, for an end-to-end "video to forces" pipeline.

References

Gilon, S., Miller, E. Y., Uhlrich, S. D. OpenCap Monocular: 3D human kinematics and musculoskeletal dynamics from a single smartphone video. arXiv preprint arXiv:2603.24733, 2026.

Keller, M., Werling, K., Shin, S., Delp, S., Pujades, S., Liu, C. K., Black, M. J. From skin to skeleton: Towards biomechanically accurate 3D digital humans. ACM Transactions on Graphics (Proc. SIGGRAPH Asia), 42(6), 2023. Code: https://github.com/MarilynKeller/SMPL2AddBiomechanics

AuthorsOpenReview ↗

[2/3] The kinematics-to-GRF estimate and the role of simulation

F̂ is a frozen estimate from the reference motion that lets the policy anticipate contact. Simulation adds what a regressor cannot, namely physically consistent forces, counterfactual analysis, robot control and joint-level quantities.

Why does the policy need the anticipated force F̂? (W2)

Thank you for this question, which showed that our notation conflated distinct quantities. §3.2 now describes them separately. Four forces appear in the method.

  • F* is the measured force-plate GRF. It enters only the reward and the critic, and only during training.
  • F̂ is a frozen estimate computed from the reference kinematics by a center-of-mass physics prior and a learned residual. The actor observes it during training and at inference, so the policy can set its gains before touchdown, before the contact force arises in the simulation.
  • The auxiliary head is a linear readout on the actor that regresses F* during training and is discarded at inference. Its equation is now in the appendix.
  • F is the force produced by the simulated contacts, the only force we evaluate.

How does F̂ compare with F and the force plates? (W2)

F̂ is closer to the force plates than the simulated force, but it only describes the recorded motion. Vertical GRF correlation with the force plates, activity means on seen / unseen subjects:

Vertical GRF r
Estimate F̂.979 / .943
GaitDynamics, kinematics-to-GRF regressor (Table 1).91 / .89
ImpedanceMimic, simulated F.80 / .76
ImpedanceMimic observing the measured GRF instead of F̂.81 / .78
ImpedanceMimic without F̂.78 / .72

If F̂ is good enough, why simulate? (W2)

For vertical force correlation on clean motion capture, a regressor is enough, which is why we treat GaitDynamics as a reference for force accuracy (§3.2). Simulation adds what a regressor cannot.

  • Physical consistency. Contact is unilateral and the center of pressure stays beneath the foot by construction. GaitDynamics places the center of pressure outside the foot on 6 % of loaded lateral-cutting and 17 % of loaded jumping frames, against at most 0.2 % for ImpedanceMimic. It also loads the ground on 30 % and 39 % of their flight frames, against 11 % and 17 % for ImpedanceMimic, whose loading there is an error in contact timing rather than a physical inconsistency (App. A.3.7).
  • Counterfactual analysis. The body, the contact model or the controller can be changed and the motion simulated again, as in our robustness tests under pelvis pushes, altered foot contact and sensory delay (Tables 23, 24).
  • Robot control. The same policies control a simulated Unitree G1 humanoid zero-shot (Table 13).
  • Joint torques and impedance. The simulation gives the joint torques and the per-joint stiffness and damping behind the forces, an interpretable representation of motor state (§4.3).
AuthorsOpenReview ↗

[3/3] Selection matrix, simulation time step, offset scale α and contact forces

S is a constant selection matrix, the 2 ms step is needed for the stability of explicit PD, α = 0.5 rad, and F is the per-foot resultant of the contact wrenches λ.

Is S the identity or an output of the policy? (W3)

Neither. S is a constant matrix S∈{0,1}m×n\mathbf{S} \in \lbrace 0,1 \rbrace^{m \times n} with one unit entry per actuated joint and zero columns for the six pelvis and ten passive coordinates. It maps the m = 15 joint torques onto the n = 31 generalized coordinates. Its definition is now in App. A.2.2 ("Human model and contact").

Why a 2 ms step rather than the 10 ms of motion capture? (Q1)

The policy and the reference already run at 10 ms. The 2 ms step is how often the PD law is re-evaluated, five times per policy step, which is what distinguishes PD control from a torque held constant over the policy step. Under explicit PD, this step is needed for stability at the gains the motions require, because the damping limit of Bound 2 scales with 1/h and the stiffness limit with 1/h² (Eq. 5). For the ankle of the reference subject (I = 0.023 kg m²), a 10 ms step would cut the admissible damping from 23.1 to 4.6 Nm s/rad, below even the cap of I/h = 11.6 Nm s/rad that Fixed-PD needs at 2 ms. Fixed-PD would then chatter in swing unless its damping were lowered further, which would worsen the trade-off with Bound 1 in contact. The soft contact model would remain admissible, as its 40 ms time constant exceeds MuJoCo's minimum of two time steps. The same limitation motivated stable PD (Tan et al., 2011), which, like implicit damping, removes Bound 2 (Table 2). A 2 ms step is also the MuJoCo default.

What is the value of α? (Q2)

α = 0.5 rad for every joint (Table 8) and for the single joint of §3.1 (Table 6). It scales the torque that the position offset can add at the reference pose, an interval of half-width αkp\alpha k^p that enters Bound 1 (Eq. 4). An earlier version of the method used α = 1.0 rad, without improvement.

How does the contact force λ differ from the simulated F? (Q3)

λ stacks the contact wrenches of the active contacts. Each of the up to 16 contact spheres in contact transmits a normal force, two friction forces, and torsional and rolling friction moments, so λ∈R6nc\boldsymbol{\lambda} \in \mathbb{R}^{6 n_c}. The per-foot GRF F is the resultant force of the contacts on that foot, and the center of pressure follows from the resultant moment of the same wrenches (App. A.2.2, "Training and implementation"). F* is the resultant force measured by the force plate, with its center of pressure. The revision now states these dimensions and the friction model (App. A.2.2, "Human model and contact").

Thank you again for these questions. We hope the clarifications address them and that you will consider raising your score. We would be glad to answer any further questions.

Reviewer_85mXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ImpedanceMimic, a physics-based human motion imitation framework that learns per-joint stiffness, damping, and position setpoints instead of relying on fixed-gain PD control. The central motivation is that accurate kinematic tracking does not necessarily imply accurate recovery of ground reaction forces (GRFs). The authors derive two simplified control constraints concerning force authority and discrete-time stability, and introduce inertia-scaled impedance parameterization for transfer across subject-specific biomechanical models. The policy is trained with kinematic imitation and measured GRF/CoP supervision, while force-plate measurements are unavailable at inference. Extensive experiments show the efficacy of the proposed method.

Critical strengths
  • The paper addresses an important limitation of physics-based motion imitation: the distinction between kinematic fidelity and kinetic fidelity, which is meaningful for biomechanics, movement analysis, and physical character control.

  • The experimental results demonstrate substantial improvements in survival and kinematics on held-out activities.

  • The single-joint analysis provides an intuitive explanation of why stiffness and damping should vary across contact and swing phases (Section 3.1, Figure 2). Importantly, the authors also examine these mechanisms in the humanoid setting (Figures 9–10, Tables 14–15). The comparison with a damping-capped Fixed-PD baseline is particularly valuable. It demonstrates that the benefits cannot be attributed solely to correcting an obviously unstable nominal controller.

  • The appendix includes extensive ablations of inertia scaling, force supervision, GRF-estimator observations, gain smoothing, and other design components (Tables 19–20). The paper also provides implementation details, evaluation splits, hyperparameters, and computational requirements. The comparison with GaitDynamics further highlights an interesting distinction between prediction accuracy and enforcing physical contact constraints.

Critical weaknesses
  • The authors interpret learned stiffness and damping trajectories as physiological signatures of human motor strategies (Section 4.3, Figure 5). However, the learned gains are optimized for a simulated PD controller and are not directly identified from human perturbation measurements. Different combinations of gains and position offsets may produce similar joint torques and motions. This concern is reinforced by the authors' own analysis in Appendix A.4.1: during walking push-off, the position-offset term accounts for 109% of the reconstructed torque, while the spring term relative to the reference contributes only 1%. The results convincingly establish that the learned gains are interpretable control-policy variables, but not necessarily that they recover physiological human joint impedance.

  • The value of fully learned, state-dependent impedance relative to simpler scheduled impedance is not completely isolated. Appendix Table 16 shows that replacing the learned gains with population-mean, activity- and phase-dependent gain schedules preserves much of the performance: average survival changes from 96.0% to 89.0%, and vertical GRF correlation from 0.79 to 0.74. Although this schedule uses oracle activity and phase information, the result suggests that relatively simple gain scheduling may explain a substantial fraction of the reported improvement. A stronger comparison against a properly trained phase-conditioned or contact-conditioned gain scheduler would clarify how much benefit comes from learned online modulation rather than structured gain variation.

Questions for authors
  • Do different random initializations or reward weights produce similar gains, or merely similar torques?

  • Given the encouraging results of Table 16, could the authors compare against a simple contact-conditioned or phase-conditioned gain scheduler, with the position-control policy trained jointly or retrained under that scheduler?

  • Could the authors evaluate moderate changes in foot-contact geometry, compliance, or friction, ideally without retraining?

2 comments
AuthorsOpenReview ↗

[1/1] Interpretation of the gains, gain scheduling, seed variation and contact robustness

Dear reviewer,

Thank you for your careful reading and for highlighting the damping-capped Fixed-PD baseline and the ablations, both of which the revision extends. Fixed-PD and ImpedanceMimic are now also compared under stable PD and implicit damping (Table 2), and all ablations use 5 seeds (Tables 27, 28). Table and section numbers refer to the revised manuscript, with submission numbers in brackets.

We have narrowed the interpretation of the gains, retrained the gain schedule you proposed, and added the seed, reward and foot-contact analyses you asked for.

Are the learned gains human impedance? (W1)

You are right that they are not. They are variables of a simulated PD controller, and kinematics and GRF alone cannot identify human joint impedance. §4.3, titled "Impedance Encodes a Physiological Signature of Task" in the submission, is now "Learned Impedance Encodes Motor State", the third contribution and the abstract use the same wording, and the section says that the gains only "partly match human ankle impedance". App. A.3.6 (submission App. A.4.1) keeps the decomposition you cite. In walking, the feedforward lever kpαδqk^p \alpha \delta q supplies 86 % of the mid-contact torque and 109 % of the push-off torque, so the push-off peak reflects this lever rather than a stiffness about the reference. The appendix also notes that a model with one torque per joint has no antagonist muscles, so kinematics and GRF cannot identify the co-activation by which humans modulate joint stiffness (Hogan, 1984).

Does a simpler gain schedule explain the benefit? (W2, Q2)

Partly, but not all of it. The submission's Table 16 swapped the gains at test time without retraining. As you suggested, the revision instead freezes the learned gains as a subject-mean, phase-indexed schedule and retrains the position policy under it (App. A.3.5, Table 20). Seen subjects, activity means over 5 seeds:

Survival (%)Joint-angle RMSE (°)Vertical GRF rLateral GRF rTorque jerk
Fixed-PD93.02.91.73.5345.3
Gain schedule, retrained88.52.85.75.6230.6
ImpedanceMimic96.92.28.80.7015.9

The schedule closes about half of the gap to ImpedanceMimic in lateral force correlation and torque jerk, but under a third in vertical force correlation and a tenth in joint-angle error. It also survives less often than fixed gains. It holds in walking and running, but fails in flexion (72.3 against 91.3 %), standing and jumping. A phase schedule therefore breaks down where the motion departs from a repeated cycle, whereas the policy adapts its gains to the state, not only to the phase.

Do different seeds and reward weights give similar gains, or only similar torques? (Q1)

Similar gains. The gain table and the principal-component analysis now use 5 seeds (Table 17, Fig. 11), like every result on the three benchmarks. To compare the gains themselves rather than the torques, we compared tables of phase-mean gains (6 activities × hip, knee and ankle × stiffness and damping × contact and swing). Between any two seeds they correlate at r ≥ .98 (mean .99), and they vary about ten times less across seeds than across cells (sd .02 against .21).

For reward weights, we compared policies trained with single reward terms removed (Tables 26, 28). Removing the amplitude normalization of the pose kernel, the end-effector term or the root term leaves the gains essentially unchanged (r ≥ .994), whereas removing the GRF reward changes them most (r = .90), so the force term shapes the gains.

Moderate changes of foot geometry, compliance and friction (Q3)

ImpedanceMimic is robust to moderate changes. We evaluated these perturbations without retraining for all four action spaces (App. A.5, Table 23). Changes relative to the unperturbed ImpedanceMimic, survival in percentage points:

PerturbationSurvivalVertical GRF r
Contact stiffness ×0.5 / ×2−0.2 / −0.1+.01 / .00
Friction ×0.6−0.5−.03
Sphere radius ×0.5−0.3.00
Sphere radius ×1.5−3.0−.21
Sphere radius ×2−25.4−.33
Coarse 4-sphere foot−0.3−.01

Moderate perturbations change survival by at most half a point. Only enlarged spheres degrade it, and they degrade every controller (at ×2, Fixed-PD loses 36.8, Torque 20.1 and Muscle 43.9 points). As smaller or fewer spheres leave every controller unaffected, the failure points to the rolling geometry of the foot.

Thank you again for questions that have made the paper stronger. We hope the revision addresses them and that you will consider raising your score. We would be glad to answer any further questions.

Thanks for the responses from the authors, which have addressed my concerns. I would like to keep my original ratings.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the revision and of the new results

Dear reviewers,

We are grateful for your careful and constructive reviews, and we truly believe your feedback has strengthened the paper, both in its experiments and in the precision of its claims. We appreciate that you consider the distinction between motion and force fidelity an important problem in physics-based motion imitation, and that you value the thorough evaluation of simulated forces against force-plate measurements across 129 participants and six activities.

Your questions converge on four themes, namely what the learned gains contribute beyond PD control and its integration scheme (85mX, H7g4), whether the baselines are compared like for like (H7g4), how far the learned gains can be read as human impedance (85mX, H7g4), and what simulation adds beyond a kinematics-to-force regressor (RKQp). We have uploaded a revised manuscript that addresses each point, and table and section numbers in our responses refer to it. The main revisions are the following:

  • Integration schemes. We retrained Fixed-PD and ImpedanceMimic under stable PD and implicit damping, each at nominal and doubled gain priors (Table 2, App. A.2.3). Both schemes remove the damping limit of explicit PD, but doubling fixed gains improves joint angles without raising the vertical-force correlation under any scheme. At matched priors, ImpedanceMimic gives the lower joint-angle error and the higher force correlations in every comparison, survives longer in all but one, and keeps the largest survival margins on withheld activities (§4.4, Tables 20, 21).
  • Existing controllers. We validated our PHC implementation on AMASS, training on a third of the original data (to fit within the discussion period), with 92.2 % test success against the reported 96.4 % (App. A.2.3, Table 10). Retrained with our full reward, PHC still tracks joint angles with 8.7 to 13.3° of error, against 2.3 to 3.2° for ImpedanceMimic (Table 2). We also implemented SimPoE as an additional baseline, which survives as often as the best controllers but tracks joint angles and vertical force less accurately. Its external residual root wrench supplies 51 to 86 % of the lateral forces, without which its survival drops sharply (Tables 2, 19–21). We also attempted to re-implement ReGAIL, which has no public code, but none of our five versions tracked the reference motions (details in H7g4's thread).
  • Torque, like for like. We retrained Torque with ImpedanceMimic's pose kernel, with the PD torque limits and at a 500 Hz policy rate (App. A.5, Table 22). None of these variants improves on the Torque of the paper, and with the same pose kernel ImpedanceMimic outperforms Torque in survival, joint angles and every force metric.
  • Additional experiments and ablations.
    • A retrained phase-indexed gain schedule closes about half of the gap to ImpedanceMimic in lateral force and torque jerk but survives less often than fixed gains, so the policy adapts its gains to the state and not only to the phase (App. A.3.5, Table 20).
    • Without the force reward, every action space loses .12 to .16 in vertical-force correlation (Table 26).
    • Under a 20 ms observation delay, matched fixed gains lose 1.6 to 8 times as much survival as the learned gains under every scheme, and moderate foot-contact changes cost at most half a point (App. A.5, Tables 23, 24).
    • Zero-shot on a simulated Unitree G1, ImpedanceMimic reaches 82.1 % survival against 56.9 % for Fixed-PD and 44.0 % for Torque, and survives longer than matched fixed gains under every scheme (§4.2, Tables 13, 14).
    • Ablations of the "ingredients" and gain ranges of ImpedanceMimic, where removing the inertia scaling multiplies torque jerk by 2.7 to 4 (Tables 25, 27, 28).
  • Broader evaluation. We added a muscle-driven action space with 80 Hill-type muscles per participant, which produces the smoothest torques but the lowest survival (§4.1, App. A.2.3, Table 1), and every result on the three benchmarks now uses 5 seeds (previously 3).
  • Narrowed claims. The learned gains are read as an encoding of motor state rather than human impedance (§4.3), the damping bound is attributed to explicit PD (§3.1), and the learned gains modulate around the prior rather than replace the choice of its scale (§4.4).
  • Clarifications and presentation. We defined the selection matrix, contact forces, force estimate, time step and α (RKQp, §3.2, App. A.2.2), and revised the text and figures for added clarity.

We are genuinely excited about motion imitation that recovers not only how humans move but also the forces behind it, and your reviews have made this work considerably stronger. If the revision addresses your concerns, we would be grateful for your support, in your scores and in helping us bring this work to the community.

We thank you again and welcome any further questions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.