Literature
Display
← All papers

Learning Variable Impedance Skills from Demonstrations with Passivity Guarantee

Yu Zhang · Long Cheng · Xiuze Xia · Haoyu Zhang

Preprint posted · arXiv ↗

Robotics5 min text · CondensedOriginal paper ↗
Contents

At a glance

This framework learns position-dependent stiffness from measured motion and interaction force, then uses Lyapunov-based conditions to regulate variable impedance. A symmetry-aware least-squares estimator can also estimate an unknown constant damping parameter. Simulations compare accuracy and speed, and a Franka robot reproduces a massage motion on a silicone dummy. This edition follows the revised manuscript, which uses the Lyapunov formulation and includes hardware experiments.

I Introduction

Copying a trajectory does not fully specify contact behavior: the robot also needs suitable compliance. The proposed pipeline estimates full stiffness matrices, learns how they vary with position, and checks conditions governing their use in an impedance controller. The contributions concern estimation, a tractable stability condition, and a demonstrated contact task.

Refer to caption
Fig. 1: The framework involves learning variable stiffness from demonstrations and implementing variable impedance control on the Franka robot to perform a massage task.

II RELATED WORK

Prior approaches obtain impedance from force–motion data, muscle signals, manually designed rules, or reinforcement learning. Changing stiffness can inject energy even when every instantaneous stiffness matrix is positive definite. Passivity-based methods address that energy flow through storage functions or virtual energy tanks. The revised paper develops a Lyapunov-based alternative and compares it with an earlier variable-impedance condition.

III Proposed Approach

The method models interaction as a virtual mass–spring–damper, estimates its parameters from demonstrations, regresses a state-dependent stiffness profile, and imposes conditions for stable reproduction.

Refer to caption
Fig. 2: The process for extracting variable stiffness parameters from demonstration data, followed by the application of these parameters in the implementation of variable impedance control, ensuring the system maintains passivity.

III-A Interaction Model

The end effector is modeled with a constant positive-definite virtual inertia. Position and velocity errors relative to a reference trajectory, measured external force, stiffness, and damping define the desired impedance relation. A slow-motion approximation simplifies the reference-acceleration term for estimation. These assumptions describe the task-space model, not an identification of every physical robot dynamic.

Equations 1–4
𝑯​𝒙¨=𝒇e+𝒇c,\boldsymbol{H}\ddot{\boldsymbol{x}}=\boldsymbol{f}^{e}+\boldsymbol{f}^{c}, (1)
𝑲t​(𝒙t−𝒙tr)+𝑫t​(𝒙˙t−𝒙˙tr)+𝑯⁡(𝒙¨t−𝒙¨tr)=𝒇te,\boldsymbol{K}_{t}(\boldsymbol{x}_{t}-\boldsymbol{x}^{r}_{t})+\boldsymbol{D}_{t}(\dot{\boldsymbol{x}}_{t}-\dot{\boldsymbol{x}}^{r}_{t})+\boldsymbol{H}(\ddot{\boldsymbol{x}}_{t}-\ddot{\boldsymbol{x}}^{r}_{t})=\boldsymbol{f}_{t}^{e}, (2)
𝑲t​𝒆t+𝑫t​𝒆˙t+𝑯​𝒆¨t=𝒇te.\boldsymbol{K}_{t}\boldsymbol{e}_{t}+\boldsymbol{D}_{t}\dot{\boldsymbol{e}}_{t}+\boldsymbol{H}\ddot{\boldsymbol{e}}_{t}=\boldsymbol{f}_{t}^{e}. (3)
𝑲t​𝒆t+𝑫t​𝒆˙t+𝑯​𝒙¨t=𝒇te.\boldsymbol{K}_{t}\boldsymbol{e}_{t}+\boldsymbol{D}_{t}\dot{\boldsymbol{e}}_{t}+\boldsymbol{H}\ddot{\boldsymbol{x}}_{t}=\boldsymbol{f}_{t}^{e}. (4)

III-B Stiffness Estimation from Demonstrations

A short sliding window treats stiffness as locally constant. Ordinary least squares estimates a full matrix and projects it to the nearest symmetric positive matrix. The proposed estimator instead expands stiffness in a symmetric basis, reducing the unknown coefficients from N² to N(N+1)/2 before the positivity correction.

With unknown constant scalar damping, the damping coefficient joins those unknowns. In the evaluated two-pass procedure, windowwise estimates determine an average damping value, then stiffness is refitted using that value. When damping is constrained to vary with stiffness through the specified eigenvalue relationship, the nonlinear coupling requires a different optimization; the paper uses CMA-ES. Neither case identifies arbitrary independent time-varying damping from unrestricted data.

Equations 5–13
𝑲t​𝒆t=𝒇te−𝑯t​𝒙¨t−𝑫t​𝒆˙t.\boldsymbol{K}_{t}\boldsymbol{e}_{t}=\boldsymbol{f}_{t}^{e}-\boldsymbol{H}_{t}\boldsymbol{\ddot{x}}_{t}-\boldsymbol{D}_{t}\boldsymbol{\dot{e}}_{t}. (5)
𝑲t​𝒙~t=𝒚~t.\boldsymbol{K}_{t}\tilde{\boldsymbol{x}}_{t}=\tilde{\boldsymbol{y}}_{t}. (6)
𝑩^t=𝒚~t​𝒙~tT​(𝒙~t​𝒙~tT)−1,\boldsymbol{\hat{B}}_{t}=\tilde{\boldsymbol{y}}_{t}\tilde{\boldsymbol{x}}^{\mathrm{T}}_{t}(\tilde{\boldsymbol{x}}_{t}\tilde{\boldsymbol{x}}^{\mathrm{T}}_{t})^{-1}, (7)
𝑩^t=𝒚~t​(𝒙~tT​𝒙~t)−1​𝒙~tT.\boldsymbol{\hat{B}}_{t}=\tilde{\boldsymbol{y}}_{t}(\tilde{\boldsymbol{x}}^{\mathrm{T}}_{t}\tilde{\boldsymbol{x}}_{t})^{-1}\tilde{\boldsymbol{x}}^{\mathrm{T}}_{t}. (8)
Bt1||Bt2|​|⋯||BtN2=Tt1||Tt2|​|⋯||TtN2​𝑷,B_{t_{1}}||B_{t_{2}}||\cdots||B_{t_{N^{2}}}=T_{t_{1}}||T_{t_{2}}||\cdots||T_{t_{N^{2}}}\boldsymbol{P}, (9)
𝑷=[(𝒙~t​𝒙~tT)−1𝟎⋯𝟎𝟎(𝒙~t​𝒙~tT)−1⋯𝟎⋱𝟎⋯𝟎(𝒙~t​𝒙~tT)−1],\boldsymbol{P}=\begin{bmatrix}(\tilde{\boldsymbol{x}}_{t}\tilde{\boldsymbol{x}}_{t}^{\mathrm{T}})^{-1}&\boldsymbol{0}&\cdots&\boldsymbol{0}\\ \boldsymbol{0}&(\tilde{\boldsymbol{x}}_{t}\tilde{\boldsymbol{x}}_{t}^{\mathrm{T}})^{-1}&\cdots&\boldsymbol{0}\\ \vdots&\vdots&\ddots&\vdots\\ \boldsymbol{0}&\cdots&\boldsymbol{0}&(\tilde{\boldsymbol{x}}_{t}\tilde{\boldsymbol{x}}_{t}^{\mathrm{T}})^{-1}\end{bmatrix}, (10)
[10⋯000⋯0⋱0⋯00],[01⋯000⋯0⋱0⋯00],⋯,[00⋯000⋯0⋱0⋯01],\begin{bmatrix}1&0&\cdots&0\\ 0&0&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&\cdots&0&0\end{bmatrix},\begin{bmatrix}0&1&\cdots&0\\ 0&0&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&\cdots&0&0\end{bmatrix},\cdots,\begin{bmatrix}0&0&\cdots&0\\ 0&0&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&\cdots&0&1\end{bmatrix},
[10⋯000⋯0⋱0⋯00],[01⋯010⋯0⋱0⋯00],⋯,[00⋯000⋯0⋱0⋯01].\begin{bmatrix}1&0&\cdots&0\\ 0&0&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&\cdots&0&0\end{bmatrix},\begin{bmatrix}0&1&\cdots&0\\ 1&0&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&\cdots&0&0\end{bmatrix},\cdots,\begin{bmatrix}0&0&\cdots&0\\ 0&0&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&\cdots&0&1\end{bmatrix}.
∑i=1N⁡(N+1)/2wit​𝑴i​𝒙~t=𝒚~t,\sum_{i=1}^{N(N+1)/2}w^{t}_{i}\boldsymbol{M}_{i}\tilde{\boldsymbol{x}}_{t}=\tilde{\boldsymbol{y}}_{t}, (11)
𝑩t=∑i=1N⁡(N+1)/2wit​𝑴i.\boldsymbol{B}_{t}=\sum_{i=1}^{N(N+1)/2}w^{t}_{i}\boldsymbol{M}_{i}. (12)
{𝑲t=𝑻tT​𝚲𝒕​𝑻t,𝑫t=𝑻tT​(ζ​𝚲𝒕𝟏𝟐)​𝑻t,\begin{cases}\boldsymbol{K}_{t}=\boldsymbol{T}_{t}^{\mathrm{T}}\boldsymbol{\Lambda_{t}}\boldsymbol{T}_{t},\\ \boldsymbol{D}_{t}=\boldsymbol{T}_{t}^{\mathrm{T}}(\zeta\boldsymbol{\Lambda^{\frac{1}{2}}_{t}})\boldsymbol{T}_{t},\\ \end{cases} (13)

III-C Stiffness Learning and Reproduction

The estimated stiffness is factorized, and Gaussian-kernel ridge regression maps position to the factor’s vectorized entries. Reassembling the matrix as LᵀL enforces positive semidefiniteness; strict positive definiteness additionally requires a nonsingular factor. This permits training from one or several demonstrations.

The reproduced stiffness shapes the task-space control force, which is mapped to joint torques with the Jacobian transpose. Learning a plausible stiffness profile alone does not guarantee stable execution; the following inequalities constrain its use together with damping and virtual inertia.

Equations 14–20
𝑳~t=Θ​(𝒔t)T​𝒘.\tilde{\boldsymbol{L}}_{t}=\Theta(\boldsymbol{s}_{t})^{\mathrm{T}}\boldsymbol{w}. (14)
J⁡(𝒘)=12​∑n=1Nd∑t=1tm{𝒘T​Θ​(𝒔t)n−(𝒙t)n}2+12​λ​𝒘T​𝒘,J(\boldsymbol{w})=\frac{1}{2}\sum_{n=1}^{N_{d}}\sum_{t=1}^{t_{m}}\begin{Bmatrix}\boldsymbol{w}^{\mathrm{T}}\Theta(\boldsymbol{s}_{t})_{n}-(\boldsymbol{x}_{t})_{n}\end{Bmatrix}^{2}+\frac{1}{2}\lambda\boldsymbol{w}^{\mathrm{T}}\boldsymbol{w}, (15)
𝒘∗=Θ​(ΘT​Θ+λ​I)−1​𝒙,\boldsymbol{w}^{*}=\Theta(\Theta^{\mathrm{T}}\Theta+\lambda I)^{-1}\boldsymbol{x}, (16)
𝑳~t=Θ​(𝒔t)T​Θ​(ΘT​Θ+λ​𝑰)−1​𝒙\tilde{\boldsymbol{L}}_{t}=\Theta(\boldsymbol{s}_{t})^{\mathrm{T}}\Theta(\Theta^{\mathrm{T}}\Theta+\lambda\boldsymbol{I})^{-1}\boldsymbol{x} (17)
𝑲​𝒆​𝒓=[k​e​r​(𝒔1,𝒔1)k​e​r​(𝒔1,𝒔2)⋯k​e​r​(𝒔1,𝒔Nd​tm)k​e​r​(𝒔2,𝒔1)k​e​r​(𝒔2,𝒔2)⋯k​e​r​(𝒔2,𝒔Nd​tm)⋱k​e​r​(𝒔Nd​tm,𝒔1)k​e​r​(𝒔Nd​tm,𝒔2)⋯k​e​r​(𝒔Nd​tm,𝒔Nd​tm)]\begin{split}\boldsymbol{Ker}=\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\\ \begin{bmatrix}ker(\boldsymbol{s}_{1},\boldsymbol{s}_{1})&ker(\boldsymbol{s}_{1},\boldsymbol{s}_{2})&\cdots&ker(\boldsymbol{s}_{1},\boldsymbol{s}_{N_{d}t_{m}})\\ ker(\boldsymbol{s}_{2},\boldsymbol{s}_{1})&ker(\boldsymbol{s}_{2},\boldsymbol{s}_{2})&\cdots&ker(\boldsymbol{s}_{2},\boldsymbol{s}_{N_{d}t_{m}})\\ \vdots&\vdots&\ddots&\vdots\\ ker(\boldsymbol{s}_{N_{d}t_{m}},\boldsymbol{s}_{1})&ker(\boldsymbol{s}_{N_{d}t_{m}},\boldsymbol{s}_{2})&\cdots&ker(\boldsymbol{s}_{N_{d}t_{m}},\boldsymbol{s}_{N_{d}t_{m}})\end{bmatrix}\end{split} (18)
𝑳~=𝒌​𝒆​𝒓∗​(𝑲​𝒆​𝒓+λ​𝑰)−1​𝒙,\tilde{\boldsymbol{L}}=\boldsymbol{ker}^{*}(\boldsymbol{Ker}+\lambda\boldsymbol{I})^{-1}\boldsymbol{x}, (19)
k​e​r​(𝒔i,𝒔j)=e(−h​‖(𝒔i−𝒔j)‖2),ker(\boldsymbol{s}_{i},\boldsymbol{s}_{j})=e^{(-h\left\|(\boldsymbol{s}_{i}-\boldsymbol{s}_{j})\right\|^{2})}, (20)

III-D Novel Lyapunov Functions

The first proposed storage function combines a shifted velocity error with stiffness-weighted position error. Sufficient matrix inequalities involve stiffness, its time derivative, damping, inertia, and a fixed positive parameter α. Unlike the comparison condition, this construction avoids the damping derivative. For the specified scalar critically damped case, the paper derives a less restrictive bound on stiffness variation.

A second storage function uses an inertia-scaled shift and an auxiliary positive-definite matrix. Its conditions retain damping-derivative terms and are compared with the earlier formulation in several dimensions. All displayed inequalities are preserved below.

The stated dissipation supply pairs external force with a shifted error output, ė + αe (or its inertia-scaled version). This is the port used by the paper’s passivity argument. It should not be silently equated with an unconditional physical force–velocity passivity guarantee for arbitrary reference motion, robot dynamics, or environment.

Equations 21–31
V⁡(𝒆,𝒆˙,t)=(𝒆˙+α​𝒆)T​𝑯​(𝒆˙+α​𝒆)2+𝒆T​𝑲t​𝒆2,V(\boldsymbol{e},\dot{\boldsymbol{e}},t)=\frac{(\dot{\boldsymbol{e}}+\alpha\boldsymbol{e})^{\mathrm{T}}\boldsymbol{H}(\dot{\boldsymbol{e}}+\alpha\boldsymbol{e})}{2}+\frac{\boldsymbol{e}^{\mathrm{T}}\boldsymbol{K}_{t}\boldsymbol{e}}{2}, (21)
V˙​(𝒆,𝒆˙,t)=(𝒆˙+α​𝒆)T​𝑭e​x​t−α​𝒆T​𝑲t​𝒆−𝒆˙T​𝑫t​𝒆˙\displaystyle\dot{V}(\boldsymbol{e},\dot{\boldsymbol{e}},t)=(\dot{\boldsymbol{e}}+\alpha\boldsymbol{e})^{\mathrm{T}}\boldsymbol{F}_{ext}-\alpha\boldsymbol{e}^{\mathrm{T}}\boldsymbol{K}_{t}\boldsymbol{e}-\boldsymbol{\dot{e}}^{\mathrm{T}}\boldsymbol{D}_{t}\boldsymbol{\dot{e}} (22)
+12​𝒆T​𝑲˙t​𝒆+α​𝒆˙T​𝑯​𝒆˙+𝒆˙T​(α2​𝑯−α​𝑫t)​𝒆\displaystyle+\frac{1}{2}\boldsymbol{e}^{\mathrm{T}}\boldsymbol{\dot{K}}_{t}\boldsymbol{e}+\alpha\boldsymbol{\dot{e}}^{\mathrm{T}}\boldsymbol{H}\boldsymbol{\dot{e}}+\boldsymbol{\dot{e}}^{\mathrm{T}}(\alpha^{2}\boldsymbol{H}-\alpha\boldsymbol{D}_{t})\boldsymbol{e}
=(𝒆˙+α​𝒆)T​𝑭e​x​t+(𝒆˙+12​α​𝒆)T​(α​𝑯−𝑫t)​(𝒆˙+12​α​𝒆)\displaystyle=(\dot{\boldsymbol{e}}+\alpha\boldsymbol{e})^{\mathrm{T}}\boldsymbol{F}_{ext}+(\boldsymbol{\dot{e}}+\frac{1}{2}\alpha\boldsymbol{e})^{\mathrm{T}}(\alpha\boldsymbol{H}-\boldsymbol{D}_{t})(\boldsymbol{\dot{e}}+\frac{1}{2}\alpha\boldsymbol{e})
+𝒆T​(12​𝑲˙t−α​𝑲t−14​α2​(α​𝑯−𝑫t))​𝒆.\displaystyle\quad+\boldsymbol{e}^{\mathrm{T}}(\frac{1}{2}\boldsymbol{\dot{K}}_{t}-\alpha\boldsymbol{K}_{t}-\frac{1}{4}\alpha^{2}(\alpha\boldsymbol{H}-\boldsymbol{D}_{t}))\boldsymbol{e}.
V⁡(t)−V⁡(0)≤∫0t(𝒆˙+α​𝒆)T​𝑭e​x​t​𝑑t,V(t)-V(0)\leq\int_{0}^{t}(\dot{\boldsymbol{e}}+\alpha\boldsymbol{e})^{\mathrm{T}}\boldsymbol{F}_{ext}dt, (23)
q¨+d⁡(t)​q˙+k⁡(t)​q=0\ddot{q}+d(t)\dot{q}+k(t)q=0 (24)
{2​ς​k⁡(t)<αk˙​(t)<2​α​k⁡(t)3k⁡(t)+2​ς​α\begin{cases}&2\varsigma\sqrt{k(t)}<\alpha\\ &\dot{k}(t)<\frac{2\alpha\sqrt{k(t)}^{3}}{\sqrt{k(t)}+2\varsigma\alpha}\end{cases} (25)
{2​ς​k⁡(t)<αk˙​(t)<2​α​k​(t)+12​α2​(α−2​ς​k⁡(t))\begin{cases}&2\varsigma\sqrt{k(t)}<\alpha\\ &\dot{k}(t)<2\alpha k(t)+\frac{1}{2}\alpha^{2}(\alpha-2\varsigma\sqrt{k(t)})\end{cases} (26)
V2​(𝒆,𝒆˙,t)=(𝒆˙+α​𝑯−1​𝒆)T​𝑯​(𝒆˙+α​𝑯−1​𝒆)2+𝒆T​𝜷t​𝒆2,V_{2}(\boldsymbol{e},\dot{\boldsymbol{e}},t)=\frac{(\dot{\boldsymbol{e}}+\alpha\boldsymbol{H}^{-1}\boldsymbol{e})^{\mathrm{T}}\boldsymbol{H}(\dot{\boldsymbol{e}}+\alpha\boldsymbol{H}^{-1}\boldsymbol{e})}{2}+\frac{\boldsymbol{e}^{\mathrm{T}}\boldsymbol{\beta}_{t}\boldsymbol{e}}{2}, (27)
V˙2​(𝒆,𝒆˙,t)=(𝒆˙+α​𝑯−1​𝒆)T​𝑭e​x​t+𝒆T​(𝟏𝟐​𝜷˙𝒕−𝜶​𝑯−𝟏​𝑲𝒕)​𝒆+𝒆˙T​(α−𝑫t)​𝒆˙+𝒆˙T​(α2​𝑯−1−𝑲t−α​𝑯−1​𝑫t+𝜷t)​𝒆˙\begin{split}\dot{V}_{2}(\boldsymbol{e},\dot{\boldsymbol{e}},t)=(\dot{\boldsymbol{e}}+\alpha\boldsymbol{H}^{-1}\boldsymbol{e})^{\mathrm{T}}\boldsymbol{F}_{ext}+\boldsymbol{e}^{\mathrm{T}}(\boldsymbol{\frac{1}{2}\dot{\beta}_{t}-\alpha\boldsymbol{H}^{-1}\boldsymbol{K}_{t}})\boldsymbol{e}\\ +\boldsymbol{\dot{e}}^{\mathrm{T}}(\alpha-\boldsymbol{D}_{t})\boldsymbol{\dot{e}}+\boldsymbol{\dot{e}}^{\mathrm{T}}(\alpha^{2}\boldsymbol{H}^{-1}-\boldsymbol{K}_{t}-\alpha\boldsymbol{H}^{-1}\boldsymbol{D}_{t}+\boldsymbol{\beta}_{t})\boldsymbol{\dot{e}}\end{split} (28)
𝑲˙t+α​𝑫˙t−2​α​𝑲t⪯𝟎,s.t.α<=min𝑡​λ¯​(𝑫t)λ¯​(𝑯)\boldsymbol{\dot{K}}_{t}+\alpha\boldsymbol{\dot{D}}_{t}-2\alpha\boldsymbol{K}_{t}\preceq\boldsymbol{0},\ \ \ s.t.\alpha<=\underset{t}{\min}{\frac{\underline{\lambda}(\boldsymbol{D}_{t})}{\overline{\lambda}(\boldsymbol{H})}} (29)
𝑲˙t+α​𝑯−1​𝑫˙t−2​α​𝑯−1​𝑲t⪯𝟎,s.t.α<=min𝑡​λ¯​(𝑫t),\boldsymbol{\dot{K}}_{t}+\alpha\boldsymbol{H}^{-1}\boldsymbol{\dot{D}}_{t}-2\alpha\boldsymbol{H}^{-1}\boldsymbol{K}_{t}\preceq\boldsymbol{0},\ \ \ s.t.\alpha<=\underset{t}{\min}{\underline{\lambda}(\boldsymbol{D}_{t})}, (30)
𝑲˙t+α​𝑫˙t/λ¯​(𝑯)−2​α​𝑲t/λ¯​(𝑯)⪯𝟎,s.t.α<=min𝑡​λ¯​(𝑫t),\boldsymbol{\dot{K}}_{t}+\alpha\boldsymbol{\dot{D}}_{t}/\overline{\lambda}(\boldsymbol{H})-2\alpha\boldsymbol{K}_{t}/\overline{\lambda}(\boldsymbol{H})\preceq\boldsymbol{0},\ \ \ s.t.\alpha<=\underset{t}{\min}{\underline{\lambda}(\boldsymbol{D}_{t})}, (31)

IV Experiment Results and Discussions

The study includes synthetic parameter-identification tests, a scalar stability comparison, and a contact experiment on a Franka Emika robot. Simulations use MATLAB; the robot implementation uses C++.

IV-A Simulation of Estimating Stiffness

Ten synthetic trajectories use a two-dimensional mass–spring–damper with a rotating stiffness ellipse. A three-sample sliding window estimates stiffness, evaluated with affine-invariant, log-Euclidean, and log-determinant matrix distances. With known damping, nearest-SPD projection is slightly more accurate in the reported case, while the proposed estimator is less sensitive to damping misspecification.

Reported per-step times are 0.23 ms for nearest-SPD estimation, 312 ms for convex optimization, and 0.24 ms for the proposed estimator. Changing the convex objective to match the least-squares normal-equation form closely aligns its estimates with the proposed method. With unknown constant damping, the two-pass estimation remains effective; convex optimization has an accuracy advantage in that example but is much slower.

Refer to caption
Fig. 3: The demonstrated stiffness ellipsoids are visualized at different time steps.
Equation 32
minimize​‖𝑲​𝑿−𝒀‖2,subject​to​𝑲⪰𝟎,\mathrm{minimize}\;\;||\boldsymbol{K}\boldsymbol{X}-\boldsymbol{Y}||_{2},\;\;\mathrm{subject\;to}\;\boldsymbol{K}\succeq\boldsymbol{0}, (32)
Refer to caption
Fig. 4: Considering a known constant damping, compare the performance of the estimated stiffness using the proposed algorithm with nearest-SPD approximation and the convex optimization.
Equation 33
minimize​‖𝑲​𝑿​𝑿T−𝒀​𝑿T‖2,subject​to​𝑲⪰𝟎,\mathrm{minimize}\;\;||\boldsymbol{K}\boldsymbol{X}\boldsymbol{X}^{\mathrm{T}}-\boldsymbol{Y}\boldsymbol{X}^{\mathrm{T}}||_{2},\;\;\mathrm{subject\;to}\;\boldsymbol{K}\succeq\boldsymbol{0}, (33)
Refer to caption
Fig. 5: With known constant damping, compare the performance of the estimated stiffness using the proposed algorithm with nearest-SPD approximation to that using convex optimization to solve the modified objective function.
Refer to caption
Fig. 6: The stiffness estimation results obtained from applying the proposed algorithm with nearest-SPD approximation and the convex optimization method while varying the damping.
Refer to caption
Fig. 7: Compare the performance of the estimated stiffness using the proposed algorithm with the convex optimization when assuming that damping is an unknown constant.
Refer to caption
Fig. 8: Compare the stability condition of the proposed Lyapunov function with the condition presented in [22].
Refer to caption
Fig. 9: Compare the energy variation of the proposed Lyapunov function with the performance described in [22].
Refer to caption
Fig. 10: Snapshots of the variable impedance control execution with the Franka Emika robot in the three-dimensional Cartesian space are presented. The robot performs a massage task on a silicone dummy, whose back is not planar. By utilizing the learned stiffness, the robot successfully completes the task.

IV-B Simulation for Validating the Novel Lyapunov Function

A one-dimensional oscillator uses sinusoidally varying desired stiffness and damping tied to its square root. In the chosen parameter setting, the proposed sufficient condition holds while the earlier comparison condition is violated at some times. The plotted storage decreases monotonically. This demonstrates a useful less-conservative case, not universal superiority for all variable-impedance systems.

Equations 34–35
m​x¨t+dt​x˙t+kt​xt=0,m\ddot{x}_{t}+d_{t}\dot{x}_{t}+k_{t}{x}_{t}=0, (34)
{xtd=10​sin⁡(0.1​t),ktd=kc+10​sin⁡(t),\begin{cases}x^{d}_{t}=10\sin(0.1t),\\ k^{d}_{t}=k_{c}+10\sin(t),\end{cases} (35)
Refer to caption
Fig. 11: The limit cycle is learned from the demonstration, with the white dotted line representing the demonstration trajectory. The black, red, and blue lines indicate trajectories that start from different random points and converge to the limit cycle.
Refer to caption
Fig. 12: The estimated stiffness is based on a segment of the demonstration trajectory, visualized using a stiffness ellipsoid.
Refer to caption
Fig. 13: Snapshots illustrate the deficiencies of using high and low stiffness for the massage task. With high stiffness (shown on the left), the massage ball may fail to conform to the surface of the silicone dummy, leading to ineffective massage. Conversely, with low stiffness (shown on the right), the robot may become trapped and require additional force to disengage, making it unsuitable for the massage task.

IV-C Validating the Robustness of the Proposed Algorithm in Robot Massage Experiments

A human guides the Franka through a massage motion on a fixed silicone dummy while motion and force are recorded. A learned limit cycle supplies reference motion, and position determines the learned stiffness. A vertical PID force-control component handles contact; damping is adjusted when the stated passivity test is not met.

High constant stiffness can lose conformity to the curved surface, while low stiffness can leave the tool trapped in a low region. The learned varying stiffness reproduces the demonstrated task more effectively in the shown trials. The same dummy and back surface are used for training and reproduction; no human-subject validation or broad quantitative generalization study is reported.

V Conclusions and Future Work

Symmetry-aware estimation and state-dependent stiffness learning offer a fast route from force–motion demonstrations to variable impedance. The practical demonstration is confined to a fixed dummy, and the theoretical conditions rely on the stated model and supply rate. Changing endpoint poses and experiments with human subjects are proposed as future work.

Concept demonstration

Open image ↗