Impedance MPC: Online Predictive Control

• Impedance MPC is a predictive control formulation that plans impedance behavior online to balance trajectory accuracy, compliance, and safety in applications like human–robot interaction and rehabilitation.
• It employs either direct optimization of stiffness and damping parameters or cost design that encodes classical impedance laws into a receding-horizon framework.
• The approach enables real-time enforcement of safety constraints, contact stability, and offset-free tracking across diverse systems such as exoskeletons, haptic steering, and dexterous manipulation.

Searching arXiv for recent and foundational papers on Impedance MPC and related formulations. Impedance Model Predictive Control (Impedance MPC) denotes a class of predictive control formulations in which impedance behavior is planned online over a receding horizon rather than fixed a priori. In the literature, this appears in two closely related forms: MPC can optimize impedance parameters directly, such as stiffness and damping, or it can optimize corrective forces or torques with a cost chosen so that the unconstrained infinite-horizon limit recovers a classical impedance law. Across haptic shared steering, compliant manipulation, physical human–robot interaction, rehabilitation exoskeletons, dexterous hands, and floating-base whole-body control, the common objective is to reconcile trajectory accuracy, compliant contact, disturbance rejection, and safety constraints within a predictive model (Izadi et al., 2020, Anand et al., 2022, Haninger et al., 2022, Cao et al., 6 Jun 2026, Cao et al., 11 Jun 2026, Cao, 12 Jun 2026, Cao, 12 Jun 2026).

1. Conceptual foundations

Classical impedance control specifies a virtual mechanical relation between tracking error and interaction force or torque. Representative forms in the cited literature include joint-space laws such as

$$\tau^\text{cmd}=\tau_\text{ff}+K_d(q_d-q)+D_d(\dot q_d-\dot q),$$

task-space laws such as

$$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$

and admittance-rendered Cartesian behavior

$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$

These formulations establish the nominal mass–spring–damper behavior that later MPC layers either reproduce or adapt online (Cao et al., 11 Jun 2026, Anand et al., 2022, Haninger et al., 2022).

The surveyed papers implement Impedance MPC in more than one structural form. In one family, the MPC decision variables are the impedance parameters themselves: future stiffness profiles $K_{t:t+T}$ in Cartesian variable impedance control, damping and stiffness vectors $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$ in haptic shared steering, or mass and damping updates in GP-based admittance control (Anand et al., 2022, Izadi et al., 2020, Haninger et al., 2022). In the other family, the predictive controller acts on an error-state model with a cost whose weights encode the desired impedance. For knee rehabilitation, fixed-base pHRI, floating-base whole-body control, and dexterous fingers, the unconstrained infinite-horizon limit recovers a classical impedance law when $Q=\operatorname{diag}(K_d,D_d)$ or $Q=\operatorname{blkdiag}(K_d,D_d)$ and the input penalty tends to zero (Cao et al., 11 Jun 2026, Cao et al., 6 Jun 2026, Cao, 12 Jun 2026, Cao, 12 Jun 2026).

A recurrent misconception is that Impedance MPC must always optimize stiffness and damping explicitly. The literature does not support that restriction. Some papers treat impedance parameters as direct decision variables, whereas others preserve the semantics of impedance through cost design, disturbance augmentation, and predictive force or torque computation. The latter class is central in recent offset-free architectures for pHRI and rehabilitation (Cao et al., 6 Jun 2026, Cao et al., 11 Jun 2026).

2. Predictive models and optimization variables

When impedance is optimized directly, the predictive model must include either the impedance state itself or a learned closed-loop map from impedance commands to future motion. In haptic shared steering, human and automation are mechanically coupled through a single steering-wheel angle $\theta_{\rm SW}$, with torques generated through backdrivable impedances

$$Z_{\rm H}(s)=B_{\rm H}s+K_{\rm H}, \qquad Z_{\rm A}(s)=B_{\rm A}s+K_{\rm A}.$$

The controller treats the automation impedance as a dynamic state,

$$\dot Z_{\rm A}(t)=\alpha_{\rm A}Z_{\rm A}(t)+\beta_{\rm A}\Gamma_{\rm A}(t),$$

and minimizes a horizon cost combining a safety term on total steering torque and a disagreement term on human–automation torque conflict. The predictive input is not torque itself but the impedance-modulation command $$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$0, and the solver is a modified least-squares scheme with post-clipping to enforce nonnegativity of stiffness and damping (Izadi et al., 2020).

In deep model predictive variable impedance control, the MPC state is the end-effector translational state

$$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$1

the control input is the Cartesian stiffness matrix $$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$2, and damping is set as $$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$3. Rather than relying on an analytic contact model, the method learns a generalized Cartesian impedance model

$$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$4

with $$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$5, using a probabilistic ensemble of neural networks and information-gain-driven exploration. The MPC cost

$$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$6

trades task performance against explicit stiffness penalization, so compliance is encoded directly in the predictive objective (Anand et al., 2022).

A third formulation couples trajectory and impedance optimization under learned uncertainty models. In model predictive impedance control with Gaussian Processes, the low-level robot is a Cartesian admittance controller, and the discrete dynamics

$$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$7

depend explicitly on the impedance parameters $$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$8. The decision vector may include both reference-force sequences and impedance updates $$M\,\delta\ddot{x}+D\,\delta\dot{x}+K\,\delta x=f_{ext},$$9, while Gaussian Process models provide mode-dependent force means and covariances. This makes impedance optimization act not only on the mean motion but also on covariance propagation, which the paper identifies as a key distinction from trajectory-only MPC (Haninger et al., 2022).

3. Impedance-equivalent MPC and disturbance-augmented architectures

A major recent line of work realizes Impedance MPC through plant reduction, disturbance augmentation, and receding-horizon quadratic programming. The central construction is to cancel known inertia, damping, gravity, or contact-consistent operational-space terms so that the residual tracking-error dynamics become a double integrator with disturbance input. For the FR3 pHRI formulation, the translational error satisfies

$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$0

with discrete model

$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$1

so $$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$2 is constant and only the input matrix varies through the operational-space inertia (Cao et al., 6 Jun 2026). For the knee exoskeleton, algebraic feedforward produces

$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$3

again yielding a constant-coefficient double integrator and a condensed QP with offline-precomputable matrices (Cao et al., 11 Jun 2026). The dexterous-finger formulation applies the same pattern to tendon-transmission dynamics, obtaining

$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$4

and a constant-$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$5 discrete model that supports 500 Hz operation with a 10-step horizon (Cao, 12 Jun 2026).

These architectures encode impedance through the MPC cost. In the knee rehabilitation system, choosing

$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$6

makes the unconstrained infinite-horizon solution recover the classical impedance structure

$$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$7

after feedforward cancellation (Cao et al., 11 Jun 2026). The fixed-base pHRI paper proves an analogous equivalence result in operational space, and the floating-base extension states an “Impedance Equivalence Theorem” under rigid contacts, fixed contact mode, $$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$8, and $$\mathbf{f}-\mathbf{f}^r=\mathbf{M}\ddot{\mathbf{x}}+\mathbf{D}\dot{\mathbf{x}}+\mathbf{K}(\mathbf{x}-\mathbf{x}_0).$$9: $K_{t:t+T}$0 There, the effective inertia is the contact-consistent operational-space inertia $K_{t:t+T}$1, so the realized impedance depends structurally on posture and contact configuration even when the cost weights are fixed (Cao, 12 Jun 2026).

The same family of methods uses augmented disturbance states to obtain offset-free tracking. In the exoskeleton, a disturbance state driven by direct SEA-based torque sensing is estimated with a Kalman filter and propagated through the horizon, yielding nominal zero steady-state error under constant or slowly varying disturbances (Cao et al., 11 Jun 2026). The FR3 pHRI architecture augments the task-space error with a random-walk disturbance state and proves offset-free steady-state tracking under asymptotically constant disturbances when the steady-state input constraint is inactive (Cao et al., 6 Jun 2026). The dexterous-hand formulation adopts the same internal-model construction with encoder-only augmented Kalman estimation, so constant contact loads are canceled in prediction rather than through integral action (Cao, 12 Jun 2026).

4. Safety constraints, contact stability, and realizability

A defining advantage of Impedance MPC over fixed-gain impedance control is that safety limits can be part of the optimization rather than post hoc saturations. In knee rehabilitation, the QP enforces hard range-of-motion, torque, and velocity constraints over the entire prediction horizon, including $K_{t:t+T}$2, $K_{t:t+T}$3 Nm, and $K_{t:t+T}$4 rad/s, explicitly tied to ISO 13482 (Cao et al., 11 Jun 2026). In the FR3 pHRI architecture, task-space force bounds are handled directly in the 30-variable QP, while joint-limit safety is enforced through a null-space inverse-barrier potential and a task-space workspace projection that modifies the effective reference near kinematic boundaries (Cao et al., 6 Jun 2026). In dexterous fingers, the QP enforces hard constraints on contact force, actuation limits, and jerk, and on the hydraulic example it adds pressure and cavitation constraints; contact-force limits are linked to ISO/TS 15066 via a 140 N contact bound (Cao, 12 Jun 2026).

For interaction with uncertain environments, safety is also expressed through probabilistic and contact-stability constraints. The GP-based formulation imposes a chance constraint on predicted force,

$K_{t:t+T}$5

so larger force uncertainty tightens the admissible mean force. It also enforces a well-damped contact condition

$K_{t:t+T}$6

where environment stiffness is estimated from the derivative of the GP mean. In the reported experiments, the well-damped constraint produced more reliable contact stabilization than the chance constraint in stiff-contact scenarios (Haninger et al., 2022).

A separate line of work identifies a structural failure mode in variable-impedance MPC for legged locomotion. When stiffness is treated as an instantaneous decision variable, the parameter-based feasible set $K_{t:t+T}$7 is strictly larger than the realizable set $K_{t:t+T}$8 under first-order actuator dynamics

$K_{t:t+T}$9

The mismatch is governed by the dimensionless parameter $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$0, and for the 1D hopping monoped the paper derives a closed-form threshold $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$1 below which no admissible stiffness command can realize the parameter-based prediction. It further derives $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$2, below which restricting the admissible stiffness range cannot repair realizability. The proposed remedy is to augment the prediction state with stiffness, which closes the mismatch by construction (Ramesh, 24 Apr 2026).

This realizability result clarifies another misconception: optimizing “stiffness profiles” is not automatically physically meaningful. If actuator bandwidth is omitted from the prediction model, the resulting controller may solve a mathematically well-posed optimization over a set of trajectories the hardware cannot traverse. In that sense, false feasibility is presented as a formulation error rather than a modeling approximation (Ramesh, 24 Apr 2026).

5. Representative applications and empirical behavior

In haptic shared steering, predictive impedance modulation is used for negotiating control authority between a human driver and an automation system physically coupled through a motorized steering wheel. The controller minimizes both deviation of total steering torque from a prescribed safety threshold and disagreement between human and automation torques. In cooperative mode, the adaptive controller drives the automation stiffness to roughly match the human stiffness; in non-cooperative mode, it reduces automation stiffness so that steering does not remain near $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$3 under conflicting torques. Compared with fixed automation impedance, the adaptive method reduces conflict and makes authority transitions smoother (Izadi et al., 2020).

In compliant manipulation, deep MPVIC evaluates learned Cartesian impedance dynamics on a Franka Emika Panda in simulation and real experiments across a Cartesian compliance task, reacting to falling objects, pushing on a table, drawer opening, and tray stabilization. The same generalized model is reused across tasks without retraining or finetuning; only the cost weights are changed. In transfer experiments, deep MPVIC required 0 additional samples to move from a model learned on one task to other tasks, whereas SAC required an additional $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$4 steps for one transfer and $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$5 for another, and PETS required $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$6 and $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$7 additional steps, respectively (Anand et al., 2022).

The GP-based framework targets semi-structured physical HRI tasks rather than a single benchmark. It learns uncertainty-aware task models from a few ($Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$8) demonstrations and is validated in contact co-manipulation, rail assembly, collaborative polishing, and a double peg-in-hole task with multiple human goals. In rail assembly, the optimized impedance increases along repeatable directions and remains low along the goal-variable direction, so the human can easily guide the rail where the target remains uncertain. In polishing, a frequency-domain disturbance-rejection term increases mass and damping in disturbance-sensitive directions while preserving low-frequency human-guided motion; reported MPC rates range from roughly 13 Hz to 26 Hz depending on the active objective and constraint set (Haninger et al., 2022).

The disturbance-augmented QP architectures report much tighter steady-state behavior than classical fixed-gain impedance control under sustained loads. On the FR3, Impedance MPC with Kalman augmentation attains sub-0.05 mm steady-state error versus 44.8 mm for classical impedance under a sustained 15 N force, and sub-millimeter tracking on four 3-D circles (Cao et al., 6 Jun 2026). On the SEA knee exoskeleton, the 500 Hz Kalman MPC is reported as offset free 0.1 mrad RMS, 0.1 mrad steady-state, and 0.2 mrad peak under 15 Nm spasm, versus a 515 mrad steady-state offset for classical impedance at the same stiffness; all MPC variants satisfy the 87 mrad clinical criterion, whereas no classical controller does (Cao et al., 11 Jun 2026). For the hydraulically actuated finger, the 500 Hz Kalman MPC achieves 0.5 mrad RMS, 0.1 mrad steady-state, and 6.6 mrad peak under 1.5 Nm contact, with realized first-move stiffness verified from 18 to 323 Nm/rad as the update rate increases (Cao, 12 Jun 2026).

In floating-base whole-body pHRI, the three-level architecture combines centroidal MPC, prioritized whole-body control, and a residual-null-space Impedance MPC for the arm end-effector. In fixed double support with an 8 N step disturbance, the full disturbance-aware controller on a 17-DOF biped reports RMS error of approximately 1.28 mm and steady-state error of approximately 0.037 mm, compared with approximately 10 mm steady-state error for PD baselines; on the Unitree G1 official MJCF, the same architecture reduces steady-state error from approximately 9.57 mm to approximately 3.90 mm despite actuator-bandwidth limits (Cao, 12 Jun 2026).

6. Boundary cases, limitations, and open directions

The literature also delineates what should not be conflated with Impedance MPC proper. A hierarchical whole-body framework for bimanual mobile manipulation uses a VLM-RAG module to output dynamic velocity limits and safety margins for a whole-body MPC while simultaneously modulating virtual stiffness and damping gains for a unified impedance-admittance controller. In that system, the MPC optimizes kinematic trajectories, whereas the impedance gains are scheduled by the semantic front-end rather than optimized within the predictive problem. The paper explicitly distinguishes this hierarchy from formulations in which $Z_{\rm A}(t)=[B_{\rm A}(t)\;K_{\rm A}(t)]^T$9 and $Q=\operatorname{diag}(K_d,D_d)$0 are decision variables of the MPC itself (Fernando et al., 21 Apr 2026).

Several limitations recur across the field. Some formulations assume that human intent and human impedance are measurable, even though practical deployment requires online estimation; the haptic steering study notes grip force as an impedance proxy but keeps intent and impedance known in simulation (Izadi et al., 2020). Learning-based impedance MPC remains sensitive to cost-weight tuning, contact discontinuities, and optimizer cost; deep MPVIC notes that CEM limits control frequency to 5–10 Hz and provides no formal safety or robustness guarantees under model uncertainty (Anand et al., 2022). GP-based impedance MPC inherits computational costs from nonlinear multiple-shooting and GP inference, and the reported contact experiments show that chance constraints can become conservative or difficult to satisfy in stiff-contact regimes (Haninger et al., 2022).

Recent disturbance-augmented QP architectures inherit strong nominal guarantees from their constant-$Q=\operatorname{diag}(K_d,D_d)$1 structure, but these guarantees depend on bounded model-reduction error, stabilizability of the reduced double-integrator plant, and disturbance models that are constant or slowly varying. The knee exoskeleton work is demonstrated on simulation and a series-elastic-actuator platform but identifies hardware implementation and clinical trials as future steps for broader validation; the floating-base extension leaves hardware validation on real humanoids to future work; the dexterous-hand formulation notes that compressibility-dominated hydraulic regimes and strong multi-DOF coupling require richer models or block estimators (Cao et al., 11 Jun 2026, Cao, 12 Jun 2026, Cao, 12 Jun 2026).

A plausible synthesis is that Impedance MPC is evolving along two technically distinct but increasingly convergent trajectories. One trajectory optimizes impedance parameters directly, often through learned or uncertainty-aware predictive models; the other encodes desired impedance behavior through cost weights, disturbance estimation, and constrained receding-horizon optimization on reduced-order plants. The false-feasibility analysis suggests that future variable-impedance MPC designs will need more careful treatment of actuator bandwidth and realizability, especially in fast locomotion and highly dynamic contact (Ramesh, 24 Apr 2026).