Literature
Display
← All papers

Sparse Control for Dynamic Movement Primitives

Patrick M. Wensing · Jean-Jacques Slotine

· IFAC-PapersOnLine ↗

Robotics4 min text · CondensedOriginal paper ↗
Contents

At a glance

A small local change to one oscillator can stop a coordinated network of rhythmic movement primitives, then allow its rhythm to resume. Contraction theory supplies conditions for this sparse inhibition, for composing primitives, and for tolerating oscillator heterogeneity. MIT Cheetah experiments demonstrate start–stop transitions in an amble gait; the analysis certifies the primitive dynamics rather than full contact and body stability.

1 Introduction

Dynamic Movement Primitives (DMPs) represent reusable attractor dynamics. The paper analyzes their interconnections through contraction and transverse contraction, then introduces spatially sparse inhibition of rhythmic DMPs. Its central distinction is between convergence to a trajectory or equilibrium and convergence to a periodic orbit up to phase. [32], [21], [33].

2 Dynamic Movement Primitives

A DMP is treated as a hierarchy of three systems: a reference filter processes commands, a canonical system defines phase, and a transformation system produces the desired movement. This common structure supports a modular stability analysis across different implementations.

2.1 Discrete (Point-To-Point) Motion Primitives

A damped spring converges to a goal while a decaying phase variable controls a learned forcing function. Because the forcing fades, it can shape the transient without changing the terminal attractor. Reference filtering smooths changes to external goal commands.

Equations 1–4
τ​y¨\displaystyle\tau\ddot{y} =k⁡(g−y)−b​y˙+f⁡(x)\displaystyle=k(g-y)-b\dot{y}+f(x) (1)
τ​x˙\displaystyle\tau\dot{x} =−αx​x\displaystyle=-\alpha_{x}x (2)
f⁡(x)=∑iΦi​(x)​wi∑iΦi​(x)​x,Φi​(x)=exp⁡(−(x−ci)22​σ12)f(x)=\frac{\sum_{i}\Phi_{i}(x)w_{i}}{\sum_{i}\Phi_{i}(x)}x,~~\Phi_{i}(x)={\rm exp}\left({-\frac{(x-c_{i})^{2}}{2\sigma_{1}^{2}}}\right) (3)
g˙=αg​(ge​x​t​(t)−g).\dot{g}=\alpha_{g}(g_{ext}(t)-g)\,. (4)

2.2 Rhythmic Motion Primitives

For rhythmic motion, an Andronov–Hopf oscillator replaces the decaying phase variable. Phase-dependent forcing shapes the output, while reference filters can smoothly adjust the oscillation center and amplitude. The canonical system retains a limit cycle rather than converging to a single point.

Equations 5–7
τ​y¨\displaystyle\tau\ddot{y} =k⁡(g−y)−b​y˙+f⁡(𝐱)\displaystyle=k(g-y)-b\dot{y}+f(\mathbf{x}) (5)
τ​x˙1\displaystyle\tau\dot{x}_{1} =ω​x2+ρ⁡(r2−x12−x22)​x1\displaystyle=\phantom{-}\omega x_{2}+\rho(r^{2}-x_{1}^{2}-x_{2}^{2})x_{1} (6)
τ​x˙2\displaystyle\tau\dot{x}_{2} =−ω​x1+ρ⁡(r2−x12−x22)​x2\displaystyle=-\omega x_{1}+\rho(r^{2}-x_{1}^{2}-x_{2}^{2})x_{2} (7)
f⁡(𝐱)=∑iΦi​(θ⁡(𝐱))​𝐰iT∑iΦi​(θ⁡(𝐱))​𝐱,Φi​(θ)=exp⁡(cos​(θ−θi)−12​σ12)\displaystyle f(\mathbf{x})=\frac{\sum_{i}\Phi_{i}(\theta(\mathbf{x}))\mathbf{w}_{i}^{T}}{\sum_{i}\Phi_{i}(\theta(\mathbf{x}))}\mathbf{x},~\Phi_{i}(\theta)={\rm exp}\left({\frac{\textrm{cos}(\theta-\theta_{i})-1}{2\sigma_{1}^{2}}}\right)\,

2.3 Commonalities

Reference and transformation subsystems are contracting in their own states. The canonical subsystem is contracting for discrete movements or transversely contracting for rhythmic movements. The combination results below propagate these properties through the hierarchy.

Equations 8–10
𝐫˙\displaystyle\dot{\mathbf{r}} =𝐟𝐫​(𝐫,𝐫e​x​t)\displaystyle=\mathbf{f}_{\mathbf{r}}(\mathbf{r},\mathbf{r}_{ext}) (Reference System) (8)
𝐱˙\displaystyle\dot{\mathbf{x}} =𝐟𝐱​(𝐱,𝐫)\displaystyle=\mathbf{f}_{\mathbf{x}}(\mathbf{x},\mathbf{r}) (Canonical System) (9)
𝐲˙\displaystyle\dot{\mathbf{y}} =𝐟𝐲​(𝐱,𝐲,𝐫)\displaystyle=\mathbf{f}_{\mathbf{y}}(\mathbf{x},\mathbf{y},\mathbf{r}) (Transformation System) (10)

3 Contraction Analysis of DMPs

The analysis establishes stability certificates for individual primitives, transformed primitives, and coupled networks.

3.1 Contraction Preliminaries

Contraction means that trajectories converge exponentially in a uniformly positive-definite metric within a forward-invariant region. Transverse contraction requires convergence only perpendicular to the flow, allowing persistent phase differences. On a compact strictly forward-invariant region without equilibria, it implies convergence to a unique limit cycle. The metric and associated generalized Jacobian express these conditions.

Equations 11–16
𝐱˙=𝐟⁡(t,𝐱).\dot{\mathbf{x}}=\mathbf{f}(t,\mathbf{x})\,. (11)
𝐀⁡(t,𝐱)=∂𝐟∂𝐱|𝐱⁡(t)\mathbf{A}(t,\mathbf{x})=\left.\frac{\partial\mathbf{f}}{\partial\mathbf{x}}\right|_{\mathbf{x}(t)} (12)
𝐌˙+𝐀T​𝐌+𝐌​𝐀≤−2​λ​𝐌\dot{\mathbf{M}}+\mathbf{A}^{T}\,\mathbf{M}+\mathbf{M}\,\mathbf{A}\leq-2\lambda\mathbf{M}
δ​𝐳​(t)=𝚯⁡(t,𝐱)​δ​𝐱​(t).\delta\mathbf{z}(t)=\boldsymbol{\Theta}(t,\mathbf{x})\,\delta\mathbf{x}(t)\,.
dd​t​δ​𝐱T​𝐌​δ​𝐱\displaystyle\frac{{\rm d}}{{\rm d}t}\delta\mathbf{x}^{T}\mathbf{M}\delta\mathbf{x} =δ​𝐱T​(𝐌˙+𝐀T​𝐌+𝐌𝐀)​δ​𝐱\displaystyle=\delta\mathbf{x}^{T}\left(\dot{\mathbf{M}}+\mathbf{A}^{T}\mathbf{M}+\mathbf{M}\mathbf{A}\right)\delta\mathbf{x} (13)
≤−2​λ​δ​𝐱T​𝐌​δ​𝐱\displaystyle\leq-2\lambda\delta\mathbf{x}^{T}\mathbf{M}\delta\mathbf{x} (14)
dd​t​δ​𝐳T​δ​𝐳=2​δ​𝐳T​𝐅s​δ​𝐳≤−2​λ​δ​𝐳T​δ​𝐳\displaystyle\frac{{\rm d}}{{\rm d}t}\delta\mathbf{z}^{T}\delta\mathbf{z}=2\delta\mathbf{z}^{T}\mathbf{F}_{s}\delta\mathbf{z}\leq-2\lambda\delta\mathbf{z}^{T}\delta\mathbf{z} (15)
δ​𝐱T​(𝐌˙+𝐀T​𝐌+𝐌​𝐀++2​λ​𝐌)​δ​𝐱≤0\delta\mathbf{x}^{T}\left(\dot{\mathbf{M}}+\mathbf{A}^{T}\,\mathbf{M}+\mathbf{M}\,\mathbf{A}++2\lambda\mathbf{M}\right)\delta\mathbf{x}\leq 0 (16)

3.2 Scaling in Space and Time

A smooth invertible spatial transformation transports the contraction metric through its inverse Jacobian. Positive time scaling likewise preserves the contraction structure under the stated uniform conditions. Rotation, translation, and scaling are special cases; more general diffeomorphisms can reshape the attractor while retaining the appropriate certificate.

Equation 17
τ⁡(t)​𝐲˙′=𝐉⁡(𝐲)​𝐟𝐲​(𝐲,𝐱,𝐫)|𝐲=𝐓−1​(𝐲′)\tau(t)\,\dot{\mathbf{y}}^{\prime}=\mathbf{J}(\mathbf{y})\,\mathbf{f}_{\mathbf{y}}(\mathbf{y},\mathbf{x},\mathbf{r})|_{\mathbf{y}=\mathbf{T}^{-1}(\mathbf{y}^{\prime})} (17)

3.3 Combination Properties

A hierarchy of contracting systems remains contracting. Combining one transversely contracting subsystem with a contracting upstream or downstream subsystem preserves transverse contraction under the stated assumptions. These results permit analyzing a DMP through its components.

3.4 Contraction Analysis of DMPs

Discrete DMP hierarchies inherit contraction; rhythmic hierarchies inherit transverse contraction for a fixed external reference. In coupled rhythmic networks, connected diffusive coupling strong enough to dominate the individual Jacobians synchronizes canonical states and yields a common stable rhythm. The displayed spectral condition is sufficient, and alternative norms may reduce conservatism.

Original figure
Figure 1: Contraction analysis readily admits stability specifications for discrete and rhythmic DMPs.
Equations 18–20
𝐫˙\displaystyle\dot{\mathbf{r}} =𝐟𝐫​(𝐫,𝐫e​x​t)\displaystyle=\mathbf{f}_{\mathbf{r}}(\mathbf{r},\mathbf{r}_{ext}) (18)
𝐱˙i\displaystyle\dot{\mathbf{x}}_{i} =𝐟𝐱​(𝐱i,𝐫)+∑j∈𝒩i𝐊i​j​(𝐱j−𝐱i)\displaystyle=\mathbf{f}_{\mathbf{x}}(\mathbf{x}_{i},\mathbf{r})+\sum_{j\in\mathcal{N}_{i}}\mathbf{K}_{ij}(\mathbf{x}_{j}-\mathbf{x}_{i}) (19)
𝐲˙i\displaystyle\dot{\mathbf{y}}_{i} =𝐟𝐲​i​(𝐲i,𝐱i,𝐫)\displaystyle=\mathbf{f}_{\mathbf{y}i}(\mathbf{y}_{i},\mathbf{x}_{i},\mathbf{r}) (20)
Original figure
Figure 2: Network of coupled rhythmic DMPs with common reference system (R), coupled canonical systems (C), and transformation systems (T).
Equation 21
λN+1​(𝐋𝐊)>maxi⁡λmax​(𝐀i​s)\lambda_{N+1}(\mathbf{L}_{\mathbf{K}})>{\max}_{i}\lambda_{\rm{max}}(\mathbf{A}_{is}) (21)

3.5 Coupled Oscillators with Multiple Frequencies

A transversely contracting network without equilibria remains so for sufficiently small continuous parameter changes. Thus nearby heterogeneous oscillators converge to a shared periodic orbit even when their natural frequencies differ. Parameter mismatch produces bounded synchronization and orbit deviations proportional to disturbance magnitude and inversely related to contraction rate; the result is local in parameter space, not unrestricted heterogeneity.

Equations 22–23
𝐱˙i=𝐟𝐱​(𝐱i,𝐫,𝝎i)+∑j∈𝒩i𝐊i​j​(𝐱j−𝐱i)\dot{\mathbf{x}}_{i}=\mathbf{f}_{\mathbf{x}}(\mathbf{x}_{i},\mathbf{r},{\boldsymbol{\omega}}_{i})+\sum_{j\in\mathcal{N}_{i}}\mathbf{K}_{ij}(\mathbf{x}_{j}-\mathbf{x}_{i}) (22)
𝐱˙i=𝐟𝐱​(𝐱i,𝐫,𝝎0)+∑j∈𝒩i𝐊i​j​(𝐱j−𝐱i)+𝐝i\dot{\mathbf{x}}_{i}=\mathbf{f}_{\mathbf{x}}(\mathbf{x}_{i},\mathbf{r},{\boldsymbol{\omega}}_{0})+\sum_{j\in\mathcal{N}_{i}}\mathbf{K}_{ij}(\mathbf{x}_{j}-\mathbf{x}_{i})+\mathbf{d}_{i} (23)
Original figure
Figure 3: Left: Uncoupled heterogeneous Van der Pol oscillators with x˙i​1=xi​2\dot{x}_{i1}=x_{i2} and x˙i​2=−ωi2​xi​1+μi​(1−xi​1)​xi​2\dot{x}_{i2}=-\omega_{i}^{2}x_{i1}+\mu_{i}(1-x_{i1})x_{i2}. Right: With diffusive coupling 𝐊i​j=diag⁡(4,4)\mathbf{K}_{ij}=\rm{diag}(4,4). New results guarantee a common period for changes in parameter heterogeneity within an open set, and continuity of the resulting periodic orbit across parameters.

4 Sparsely-Inhibited Rhythmic DMPs

Sparse inhibition changes a rhythmic network into a contracting one by modifying only a small portion of its canonical dynamics.

4.1 Local Influence of Contracting Dynamics

When a transverse-contraction region overlaps a forward-invariant contraction region, trajectories in their union converge to the unique equilibrium in the overlap. A sufficiently strong contracting vector field can dominate another smooth field on a compact region. These facts enable local capture of an orbit and motivate sequential composition.

Equations
𝐌˙+𝐀T​𝐌+𝐌𝐀\displaystyle\dot{\mathbf{M}}+\mathbf{A}^{T}\mathbf{M}+\mathbf{M}\mathbf{A} ≤∂𝐌∂𝐱⋅𝐟1+𝐀1T​𝐌+𝐀1​𝐌−2​α​λ2​𝐌\displaystyle\leq\frac{\partial\mathbf{M}}{\partial\mathbf{x}}\cdot\mathbf{f}_{1}+\mathbf{A}_{1}^{T}\mathbf{M}+\mathbf{A}_{1}\mathbf{M}-2\alpha\lambda_{2}\mathbf{M}\,
β=inf{b∈ℝ|∀𝐱∈𝒞,∂𝐌∂𝐱⋅𝐟1+𝐀1T𝐌+𝐀1𝐌<b𝐌}\beta=\inf\left\{b\in\mathbb{R}~{\Large|}~\forall{\mathbf{x}}\in\mathcal{C},~\frac{\partial\mathbf{M}}{\partial\mathbf{x}}\cdot\mathbf{f}_{1}+\mathbf{A}_{1}^{T}\mathbf{M}+\mathbf{A}_{1}\mathbf{M}<b\mathbf{M}\right\}

4.2 Application to Sparse Inhibition of Rhythmic DMPs

Diffusive coupling plus a contracting influence at one node creates a contracting network contribution. If coupling and inhibition are strong enough, that contribution suppresses the entire network’s rhythm. Inhibition can activate when the controlled oscillator enters a selected region. A directed-network extension requires every node to be reachable from the inhibited node.

Equations 24
dd​t​𝐱=𝐅𝐱​(𝐱,𝐫)−𝐋​𝐱\frac{{\rm d}}{{\rm d}t}\mathbf{x}=\mathbf{F}_{\mathbf{x}}(\mathbf{x},\mathbf{r})-\mathbf{L}\,\mathbf{x}
dd​t​𝐱=𝐅𝐱​(𝐱,𝐫)−𝐋​𝐱+[𝐠​(𝐱1)T,𝟎,…,𝟎]T\frac{{\rm d}}{{\rm d}t}\mathbf{x}=\mathbf{F}_{\mathbf{x}}(\mathbf{x},\mathbf{r})-\mathbf{L}\,\mathbf{x}+[\mathbf{g}(\mathbf{x}_{1})^{T},\mathbf{0},\ldots,\mathbf{0}]^{T}\ (24)
δ​𝐱T​∂𝐟inh∂𝐱​δ​𝐱=−δ​𝐱T​𝐋𝐊​δ​𝐱+δ​𝐱1T​(∂𝐠∂𝐱1+∂𝐠∂𝐱1T)​δ​𝐱1\displaystyle\delta\mathbf{x}^{T}\frac{\partial\mathbf{f}_{\rm inh}}{\partial\mathbf{x}}\delta\mathbf{x}=-\delta\mathbf{x}^{T}\mathbf{L}_{\mathbf{K}}\delta\mathbf{x}+\delta\mathbf{x}^{T}_{1}\left(\frac{\partial\mathbf{g}}{\partial\mathbf{x}_{1}}+\frac{\partial\mathbf{g}}{\partial\mathbf{x}_{1}}^{T}\right)\delta\mathbf{\mathbf{x}}_{1}

5 Experiments With the MIT Cheetah

Four coupled oscillators coordinate a four-beat amble on MIT Cheetah, with filtered speed and turning commands. Each leg’s transformation system drives three joints. In the illustrated run, inhibition activates at 108.3 seconds and is removed at 112.4 seconds, stopping and restoring the coordinated oscillations. Body-feedback torques assist balance but are outside the primitive-level proof.

Refer to caption
Figure 4: SI-DMPs are applied to provide rhythmic oscillations for a four-beat amble gait.
Equations 25–28
v˙\displaystyle\dot{v} =αr​(ve​x​t−v)\displaystyle=\alpha_{r}(v_{ext}-v) (25)
γ˙\displaystyle\dot{\gamma} =αr​(γe​x​t−γ)\displaystyle=\alpha_{r}(\gamma_{ext}-\gamma) (26)
𝐱˙i=𝐟𝐱​(𝐱i,𝐫)+∑j∈𝒩i𝐊i​j​(𝐑⁡(ϕi​j)​𝐱j−𝐱i)\displaystyle\dot{\mathbf{x}}_{i}=\mathbf{f}_{\mathbf{x}}(\mathbf{x}_{i},\mathbf{r})+\sum_{j\in\mathcal{N}_{i}}\mathbf{K}_{ij}(\mathbf{R}(\phi_{ij})\mathbf{x}_{j}-\mathbf{x}_{i}) (27)
θ¨im=km​(gθm​(𝐱i,𝐫)−θim)+bm​(gθ˙m​(𝐱i,𝐫)−θ˙im)\ddot{\theta}_{i}^{m}=k^{m}(g_{\theta}^{m}(\mathbf{x}_{i},\mathbf{r})-\theta_{i}^{m})+b^{m}(g_{\dot{\theta}}^{m}(\mathbf{x}_{i},\mathbf{r})-\dot{\theta}_{i}^{m}) (28)
Original figure
Figure 5: Target kinematics for hip and knee joints during an amble gait with Sparsely-Inhibited Rhythmic DMPs.
Original figure
Original figure
Figure 6: Andronov-Hopf oscillator with and without switched inhibition. Upon a state entering a region 𝒞\mathcal{C}, a strong contracting dynamic is switched, rendering 𝒞\mathcal{C} forward invariant and contracting.
Original figure
Figure 7: Canonical states 𝐱i\mathbf{x}_{i} across the addition and removal of sparse inhibition to the DMP network.

6 Discussion

The framework suggests contact-triggered synchronization, distributed inhibitory voting, and hierarchical control. The reported experiments lack contact sensing, and the analysis does not model how limb forces stabilize the body. Robust locomotion on uncertain terrain therefore remains an extension rather than a demonstrated guarantee.

Equation
𝐟inh​(𝐱)=−𝐋​𝐱+[α1​𝐠1​(𝐱1)T,…,αN​𝐠N​(𝐱N)T]\mathbf{f}_{\rm inh}(\mathbf{x})=-\mathbf{L}\,\mathbf{x}+[\alpha_{1}\mathbf{g}_{1}(\mathbf{x}_{1})^{T},\ldots,\alpha_{N}\mathbf{g}_{N}(\mathbf{x}_{N})^{T}]

Appendix A Selected Proofs

The appendix supplies the metric construction and disturbance bound used in the main analysis.

A.1 Proof of Theorem 1

Starting from a rank-deficient metric that contracts transverse directions, the proof separates tangent and transverse differential coordinates. Integral solutions cancel their cross terms and construct a positive-definite metric. The resulting generalized Jacobian has one zero eigenvalue and strictly negative remaining eigenvalues in its symmetric part.

Equations 29–44
𝐌˙s+𝐀T​𝐌s+𝐌s​𝐀≤−2​λ​𝐌s.\dot{\mathbf{M}}_{s}+\mathbf{A}^{T}\mathbf{M}_{s}+\mathbf{M}_{s}\mathbf{A}\leq-2\lambda\mathbf{M}_{s}\,. (29)
𝐌s=∫0∞𝐕​(t,𝐱)T​𝐐​(𝐱⁡(t))​𝐕​(t,𝐱)​𝑑t\mathbf{M}_{s}=\int_{0}^{\infty}\mathbf{V}(t,\mathbf{x})^{T}\mathbf{Q}(\mathbf{x}(t))\mathbf{V}(t,\mathbf{x}){\rm d}t
𝐯˙=(𝐀−𝐟𝐟T𝐟T​𝐟​(𝐀+𝐀T))⏟𝐀𝐯​𝐯.\dot{\mathbf{v}}=\underbrace{\left(\mathbf{A}-\frac{\mathbf{f}\mathbf{f}^{T}}{\mathbf{f}^{T}\mathbf{f}}(\mathbf{A}+\mathbf{A}^{T})\right)}_{\mathbf{A}_{\mathbf{v}}}\mathbf{v}\,.
𝐌s=𝚷T​(𝐱)​𝐌~s​(𝐱)​𝚷​(𝐱)\mathbf{M}_{s}=\boldsymbol{\Pi}^{T}(\mathbf{x})\,\tilde{\mathbf{M}}_{s}(\mathbf{x})\,\boldsymbol{\Pi}(\mathbf{x})
δ​𝐱˙=𝐀⁡(t,𝐱)​δ​𝐱\dot{\delta\mathbf{x}}=\mathbf{A}(t,\mathbf{x})\delta\mathbf{x}
[δ​𝐳1δ​𝐳2]=[𝐟T|𝐟|2𝐌~s1/2​𝚷]⏟𝚯𝐱​δ​𝐱\begin{bmatrix}\delta\mathbf{z}_{1}\\ \delta\mathbf{z}_{2}\end{bmatrix}=\underbrace{\begin{bmatrix}\frac{\mathbf{f}^{T}}{|\mathbf{f}|^{2}}\\[8.61108pt] \tilde{\mathbf{M}}_{s}^{1/2}\boldsymbol{\Pi}\end{bmatrix}}_{\boldsymbol{\Theta}_{\mathbf{x}}}\delta\mathbf{x}
dd​t​[δ​𝐳1δ​𝐳2]=[0𝐀12​(t,𝐱)0𝐀22​(t,𝐱)]​[δ​𝐳1δ​𝐳2]\frac{{\rm d}}{{\rm d}t}\begin{bmatrix}\delta\mathbf{z}_{1}\\ \delta\mathbf{z}_{2}\end{bmatrix}=\begin{bmatrix}0&\mathbf{A}_{12}(t,\mathbf{x})\\ 0&\mathbf{A}_{22}(t,\mathbf{x})\end{bmatrix}\begin{bmatrix}\delta\mathbf{z}_{1}\\ \delta\mathbf{z}_{2}\end{bmatrix}
dd​t​δ​𝐳2T​δ​𝐳2\displaystyle\frac{{\rm d}}{{\rm d}t}\delta\mathbf{z}_{2}^{T}\,\delta\mathbf{z}_{2} =dd​t​δ​𝐱T​𝐌s​δ​𝐱\displaystyle=\frac{{\rm d}}{{\rm d}t}\delta\mathbf{x}^{T}\mathbf{M}_{s}\delta\mathbf{x} (30)
=δ​𝐱T​(𝐌˙s+𝐀T​𝐌s+𝐌s​𝐀)​δ​x\displaystyle=\delta\mathbf{x}^{T}\left(\dot{\mathbf{M}}_{s}+\mathbf{A}^{T}\mathbf{M}_{s}+\mathbf{M}_{s}\mathbf{A}\right)\delta x (31)
≤−2​λ​δ​𝐱T​𝐌s​δ​𝐱\displaystyle\leq-2\lambda\delta\mathbf{x}^{T}\mathbf{M}_{s}\delta\mathbf{x} (32)
=−2​λ​δ​𝐳2T​δ​𝐳2.\displaystyle=-2\lambda\delta\mathbf{z}_{2}^{T}\,\delta\mathbf{z}_{2}\,. (33)
𝐌z​(𝐱)=[1𝐌21T​(𝐱)𝐌21​(𝐱)𝐌22​(𝐱)].\mathbf{M}_{z}(\mathbf{x})=\begin{bmatrix}1&\mathbf{M}_{21}^{T}(\mathbf{x})\\ \mathbf{M}_{21}(\mathbf{x})&\mathbf{M}_{22}(\mathbf{x})\end{bmatrix}\,.
dd​t​(δ​𝐳T​𝐌z​δ​𝐳)\displaystyle\frac{{\rm d}}{{\rm d}t}\left(\delta\mathbf{z}^{T}\mathbf{M}_{z}\delta\mathbf{z}\right) =dd​t​(δ​𝐳1T​δ​𝐳1)+dd​t​(δ​𝐳2T​𝐌22​δ​𝐳2)\displaystyle=\frac{{\rm d}}{{\rm d}t}\left(\delta\mathbf{z}_{1}^{T}\,\delta\mathbf{z}_{1}\right)+\frac{{\rm d}}{{\rm d}t}\left(\delta\mathbf{z}_{2}^{T}\,\mathbf{M}_{22}\delta\mathbf{z}_{2}\right)
+dd​t​2​(δ​𝐳2T​𝐌21​δ​𝐳1)\displaystyle\phantom{=}+\frac{{\rm d}}{{\rm d}t}2\left(\delta\mathbf{z}_{2}^{T}\mathbf{M}_{21}\delta\mathbf{z}_{1}\right)
=δ​𝐳2T​(𝐌˙22+𝐌22​𝐀22+𝐀22T​𝐌22\displaystyle=\delta\mathbf{z}_{2}^{T}\left(\dot{\mathbf{M}}_{22}+\mathbf{M}_{22}\mathbf{A}_{22}+\mathbf{A}_{22}^{T}\mathbf{M}_{22}\right.
+𝐌21​𝐀12+𝐀12T​𝐌21T)​δ​𝐳2\displaystyle\phantom{=}\quad\quad\quad~~\left.+\mathbf{M}_{21}\mathbf{A}_{12}+\mathbf{A}_{12}^{T}\mathbf{M}_{21}^{T}\right)\,\delta\mathbf{z}_{2}
+2​δ​𝐳2T​(𝐌˙21+𝐀22T​𝐌21+𝐀12T)​δ​𝐳1\displaystyle\quad+2\delta\mathbf{z}_{2}^{T}\left(\dot{\mathbf{M}}_{21}+\mathbf{A}_{22}^{T}\mathbf{M}_{21}+\mathbf{A}_{12}^{T}\right)\delta\mathbf{z}_{1}
𝐌˙21+𝐀22T​𝐌21+𝐀12T=0.\dot{\mathbf{M}}_{21}+\mathbf{A}_{22}^{T}\mathbf{M}_{21}+\mathbf{A}_{12}^{T}=0\,. (34)
𝐌21​(𝐱)=∫0∞𝐔2T​(τ,𝐱)​𝐀12T​(τ,𝐱)​𝑑τ\mathbf{M}_{21}(\mathbf{x})=\int_{0}^{\infty}\mathbf{U}_{2}^{T}(\tau,\mathbf{x})\mathbf{A}_{12}^{T}(\tau,\mathbf{x}){\rm d}\tau
dd​t​𝐔2​(t,𝐱)=𝐀22​(t,𝐱)​𝐔2​(t,𝐱)𝐔2​(0,𝐱)=𝐈\frac{{\rm d}}{{\rm d}t}\mathbf{U}_{2}(t,\mathbf{x})=\mathbf{A}_{22}(t,\mathbf{x})\,\mathbf{U}_{2}(t,\mathbf{x})\quad\mathbf{U}_{2}(0,\mathbf{x})=\mathbf{I}
𝐔2T​(t,𝐱)​(𝐀22T​(t,𝐱)​𝐌21​(t,𝐱)+𝐌˙21​(t,𝐱))\displaystyle\mathbf{U}_{2}^{T}(t,\mathbf{x})\left(\mathbf{A}_{22}^{T}(t,\mathbf{x})\mathbf{M}_{21}(t,\mathbf{x})+\dot{\mathbf{M}}_{21}(t,\mathbf{x})\right) (35)
=\displaystyle= −𝐔2T​(t,𝐱)​𝐀12T​(t,𝐱)\displaystyle-\mathbf{U}_{2}^{T}(t,\mathbf{x})\mathbf{A}_{12}^{T}(t,\mathbf{x}) (36)
dd​t​(𝐔2T​(t,𝐱)​𝐌21​(t,𝐱))=−𝐔2T​(t,𝐱)​𝐀12T​(t,𝐱).\frac{{\rm d}}{{\rm d}t}\left(\mathbf{U}_{2}^{T}(t,\mathbf{x})\,\mathbf{M}_{21}(t,\mathbf{x})\right)=-\mathbf{U}_{2}^{T}(t,\mathbf{x})\mathbf{A}_{12}^{T}(t,\mathbf{x})\,.
−𝐔2T(0,𝐱)𝐌21(0,𝐱)=∫0∞−𝐔2T(t,𝐱)𝐀12T(t,𝐱)dt-\mathbf{U}_{2}^{T}(0,\mathbf{x})\,\mathbf{M}_{21}(0,\mathbf{x})=\int_{0}^{\infty}-\mathbf{U}_{2}^{T}(t,\mathbf{x})\,\mathbf{A}_{12}^{T}(t,\mathbf{x}){\rm d}t
dd​t​(δ​𝐳T​𝐌z​δ​𝐳)\displaystyle\frac{{\rm d}}{{\rm d}t}\left(\delta\mathbf{z}^{T}\mathbf{M}_{z}\delta\mathbf{z}\right) =δ​𝐳2T​(𝐌˙22+𝐌22​𝐀22+𝐀22T​𝐌22+\displaystyle=\delta\mathbf{z}_{2}^{T}\left(\dot{\mathbf{M}}_{22}+\mathbf{M}_{22}\mathbf{A}_{22}+\mathbf{A}_{22}^{T}\mathbf{M}_{22}+\right.
𝐌21​𝐀12+𝐀12T​𝐌21T)​δ​𝐳2\displaystyle\phantom{=}\left.~~~~~~~~~\mathbf{M}_{21}\mathbf{A}_{12}+\mathbf{A}_{12}^{T}\mathbf{M}_{21}^{T}\right)\,\delta\mathbf{z}_{2} (37)
𝐌˙22+𝐌22​𝐀22+𝐀22T​𝐌22+2​r​𝐌22\displaystyle\dot{\mathbf{M}}_{22}+\mathbf{M}_{22}\mathbf{A}_{22}+\mathbf{A}_{22}^{T}\mathbf{M}_{22}+2r\mathbf{M}_{22}~
+𝐌21​𝐀12+𝐀12T​𝐌21T+𝐐\displaystyle+\mathbf{M}_{21}\mathbf{A}_{12}+\mathbf{A}_{12}^{T}\mathbf{M}_{21}^{T}+\mathbf{Q} =0\displaystyle=0 (38)
𝐌22​(𝐱)=∫0∞e2​r​t​𝐔2T​(𝐌21​𝐀12+𝐀12T​𝐌21T+𝐐)​𝐔2​𝑑t\mathbf{M}_{22}(\mathbf{x})=\int_{0}^{\infty}{\rm e}^{2rt}\mathbf{U}_{2}^{T}\left(\mathbf{M}_{21}\mathbf{A}_{12}+\mathbf{A}_{12}^{T}\mathbf{M}_{21}^{T}+\mathbf{Q}\right)\mathbf{U}_{2}\,{\rm d}t
dd​t​(e2​r​t​𝐔2​(t,𝐱)T​𝐌22​(t,𝐱)​𝐔2​(t,𝐱))\displaystyle\frac{{\rm d}}{{\rm d}t}\left({\rm e}^{2rt}\mathbf{U}_{2}(t,\mathbf{x})^{T}\mathbf{M}_{22}(t,\mathbf{x})\mathbf{U}_{2}(t,\mathbf{x})\right) (39)
=−e2​r​t​𝐔2​(t,𝐱)T​(𝐌21​𝐀12+𝐀12T​𝐌21T+𝐐)​𝐔2​(t,𝐱)\displaystyle~=-{\rm e}^{2rt}\mathbf{U}_{2}(t,\mathbf{x})^{T}\left(\mathbf{M}_{21}\mathbf{A}_{12}+\mathbf{A}_{12}^{T}\mathbf{M}_{21}^{T}+\mathbf{Q}\right)\mathbf{U}_{2}(t,\mathbf{x}) (40)
dd​t​(δ​𝐳T​𝐌z​δ​𝐳)=−δ​𝐳2T​(𝐐+2​r​𝐌22)​δ​𝐳2\displaystyle\frac{{\rm d}}{{\rm d}t}\left(\delta\mathbf{z}^{T}\mathbf{M}_{z}\delta\mathbf{z}\right)=-\delta\mathbf{z}_{2}^{T}\left(\mathbf{Q}+2r\mathbf{M}_{22}\right)\,\delta\mathbf{z}_{2} (41)
dd​t​(δ​𝐱T​𝚯T​𝚯​δ​𝐱)=\displaystyle\frac{{\rm d}}{{\rm d}t}\left(\delta\mathbf{x}^{T}\boldsymbol{\Theta}^{T}\boldsymbol{\Theta}\delta\mathbf{x}\right)= dd​t​(δ​𝐲T​δ​𝐲)\displaystyle\frac{{\rm d}}{{\rm d}t}\left(\delta\mathbf{y}^{T}\delta\mathbf{y}\right) (42)
=\displaystyle= 2​δ​𝐲T​𝐅s​δ​𝐲\displaystyle 2\delta\mathbf{y}^{T}\mathbf{F}_{s}\delta\mathbf{y} (43)
=\displaystyle= −δ​𝐳2T​(𝐐+2​r​𝐌22)​δ​𝐳2\displaystyle-\delta\mathbf{z}_{2}^{T}\left(\mathbf{Q}+2r\mathbf{M}_{22}\right)\,\delta\mathbf{z}_{2} (44)

A.2 Disturbed Transverse Contracting Systems

A virtual time-scaled trajectory aligns the nominal flow with a disturbed trajectory. Bounding the derivative of Riemannian distance gives an orbit-tube radius equal to the disturbance bound divided by contraction rate, multiplied by a metric-conditioning factor. The disturbed trajectory must remain in the compact contraction region.

Equations 45–47
𝐱˙d=𝐟⁡(𝐱d)+𝐰⁡(t)\dot{\mathbf{x}}_{d}=\mathbf{f}(\mathbf{x}_{d})+\mathbf{w}(t)\,
𝐲˙=𝐟⁡(𝐲)​u\dot{\mathbf{y}}=\mathbf{f}(\mathbf{y})u (45)
𝜸(𝐱1,𝐱2):{𝜸(⋅)∈𝒞∞([0,1],𝒦) s.t. ∂∂s𝜸(s)≠0\displaystyle\boldsymbol{\gamma}(\mathbf{x}_{1},\mathbf{x}_{2}):\{\boldsymbol{\gamma}(\cdot)\in\mathcal{C}^{\infty}([0,1],\mathcal{K})\textrm{~s.t.~}\frac{\partial}{\partial s}\boldsymbol{\gamma}(s)\neq 0
∀s∈(0,1), 𝜸(0)=𝐱1,𝜸(1)=𝐱2}\displaystyle\quad\quad\quad\quad~~~~~~~~~~\forall s\in(0,1)\textrm{, }\boldsymbol{\gamma}(0)=\mathbf{x}_{1},~\boldsymbol{\gamma}(1)=\mathbf{x}_{2}\}
d⁡(𝐱1,𝐱2):=infγ∈𝜸⁡(𝐱1,𝐱2)∫01𝜸s​(s)T​𝐌​(𝜸⁡(s))​𝜸s​(s)​d​sd(\mathbf{x}_{1},\mathbf{x}_{2}):=\inf_{\gamma\in\boldsymbol{\gamma}(\mathbf{x}_{1},\mathbf{x}_{2})}\int_{0}^{1}\sqrt{\boldsymbol{\gamma}_{s}(s)^{T}\mathbf{M}(\boldsymbol{\gamma}(s))\boldsymbol{\gamma}_{s}(s)}\textrm{d}s
e⁡(𝐱1,𝐱2):=infγ∈𝜸⁡(𝐱1,𝐱2)∫01𝜸s​(s)T​𝐌​(𝜸⁡(s))​𝜸s​(s)​d​se(\mathbf{x}_{1},\mathbf{x}_{2}):=\inf_{\gamma\in\boldsymbol{\gamma}(\mathbf{x}_{1},\mathbf{x}_{2})}\int_{0}^{1}\boldsymbol{\gamma}_{s}(s)^{T}\mathbf{M}(\boldsymbol{\gamma}(s))\boldsymbol{\gamma}_{s}(s)\textrm{d}s
12​D+​e​(𝐱d​(t),𝐲⁡(t))=\displaystyle\frac{1}{2}D^{+}e(\mathbf{x}_{d}(t),\mathbf{y}(t))= 𝜸s​(0)T​𝐌​(𝐱d)​(𝐟⁡(𝐱d)+𝐰)\displaystyle\boldsymbol{\gamma}_{s}(0)^{T}\mathbf{M}(\mathbf{x}_{d})(\mathbf{f}(\mathbf{x}_{d})+\mathbf{w})
−𝜸s​(1)T​𝐌​(𝐲)​𝐟​(𝐲)​u\displaystyle-\boldsymbol{\gamma}_{s}(1)^{T}\mathbf{M}(\mathbf{y})\mathbf{f}(\mathbf{y})u (46)
𝜸s​(0)T​𝐌​(𝐱d)​𝐟​(𝐱d)−𝜸s​(1)T​𝐌​(𝐲)​𝐟​(𝐲)​u<−λ​e​(𝐱d,𝐲)\boldsymbol{\gamma}_{s}(0)^{T}\mathbf{M}(\mathbf{x}_{d})\mathbf{f}(\mathbf{x}_{d})-\boldsymbol{\gamma}_{s}(1)^{T}\mathbf{M}(\mathbf{y})\mathbf{f}(\mathbf{y})u<-\lambda e(\mathbf{x}_{d},\mathbf{y})
12​D+​e​(𝐱d​(t),𝐲⁡(t))≤\displaystyle\frac{1}{2}D^{+}e(\mathbf{x}_{d}(t),\mathbf{y}(t))\leq d⁡(𝐱d​(t),𝐲⁡(t))​|𝚯⁡(𝐱d​(t))​𝐰​(t)|2\displaystyle d(\mathbf{x}_{d}(t),\mathbf{y}(t))|\boldsymbol{\Theta}(\mathbf{x}_{d}(t))\mathbf{w}(t)|_{2}
−λ​e​(𝐱d​(t),𝐲⁡(t))\displaystyle-\lambda e(\mathbf{x}_{d}(t),\mathbf{y}(t)) (47)
Concept demonstration

Open image ↗