Literature
Display
← All papers

Composable Geometric Motion Policies using Multi-Task Pullback Bundle Dynamical Systems

Andrew Bylard · Riccardo Bonalli · Marco Pavone

· IEEE International Conference on Robotics and Automation (ICRA 2021) ↗

Robotics7 min text · CondensedOriginal paper ↗
Contents

At a glance

PBDS combines task-space motion rules into one acceleration policy while respecting non-Euclidean geometry. It separates metrics that define task behavior from weights that prioritize tasks, permits velocity-dependent priorities, and encodes constraints through metric curvature. The combination reduces to weighted least squares. Sphere examples and a Franka mug-grasping demonstration illustrate the approach; the reported implementation runs at 300–500 Hz.

I Introduction

Reactive controllers must combine attraction, damping, collision avoidance, and other goals on spaces such as joint configurations, rotations, or a sphere. Repulsive potentials can create unwanted minima, while a coordinate-dependent geometric construction can produce different physical behavior under a change of chart.

PBDS pulls task dynamics back to the robot through well-defined bundle constructions. Its key design choice is to separate a position-dependent behavior metric from a position/velocity-dependent task-weighting pseudometric. This lets the designer change priority without also changing the intrinsic task dynamics. The paper compares this with the specific GDS/RMP formulation of RMPflow. [17]

Refer to caption
Fig. 1: Example tree of PBDS task mappings designed to move a ball along the surface of a sphere to a goal while avoiding obstacles. Depicted are manifolds representing: (black) joint configuration for a 7-DoF robot arm with two fully revolute joints, (purple) ball pose, (red) distances to obstacles, (blue) position on the sphere, and (green) distance to the goal. Note that damping can be defined on 𝕋2×ℝ5\mathbb{T}^{2}\times\mathbb{R}^{5}, S​E​(3)SE(3), and/or 𝕊2\mathbb{S}^{2} as desired.

II Geometric Preliminaries

A manifold represents configurations, its tangent bundle represents velocities, and its cotangent bundle naturally represents forces. A Riemannian metric defines lengths and converts forces to accelerations; a connection adds the curvature terms needed to define acceleration independently of coordinates. Local equations use Christoffel symbols, but the resulting objects must agree when charts change.

Equations 1
σ¨k​(t)+σ˙i​(t)​σ˙j​(t)​Γi​jk​(σ⁡(t))=gk​j​(σ⁡(t))​ℱj​(σ⁡(t),σ˙​(t)),\ddot{\sigma}^{k}(t)+\dot{\sigma}^{i}(t)\dot{\sigma}^{j}(t)\Gamma^{k}_{ij}(\sigma(t))=g^{kj}(\sigma(t))\mathcal{F}_{j}(\sigma(t),\dot{\sigma}(t)),\hskip-1.0pt (1)
𝝈¨​(t)+𝚵⁡(𝝈⁡(t),𝝈˙​(t))=𝒈−1​(𝝈⁡(t))​𝓕​(𝝈⁡(t),𝝈˙​(t)).\ddot{\bm{\sigma}}(t)+\bm{\Xi}(\bm{\sigma}(t),\dot{\bm{\sigma}}(t))=\bm{g}^{-1}(\bm{\sigma}(t))\bm{\mathcal{F}}(\bm{\sigma}(t),\dot{\bm{\sigma}}(t)).
𝝈¨+𝚵=𝒈−1​𝓕.\ddot{\bm{\sigma}}+\bm{\Xi}=\bm{g}^{-1}\bm{\mathcal{F}}.
T​M{\lx@inpgf@ignorespaces TM}f∗​T​N{\lx@inpgf@ignorespaces f^{*}TN}T​N{\lx@inpgf@ignorespaces TN}M{\lx@inpgf@ignorespaces M}N{\lx@inpgf@ignorespaces N}f∗​d​f{\lx@inpgf@ignorespaces f^{*}df}πM{\lx@inpgf@ignorespaces\pi_{M}}f∗​πN{\lx@inpgf@ignorespaces f^{*}\pi_{N}}f∗​IdT​N{\lx@inpgf@ignorespaces f^{*}\textrm{Id}_{TN}}πN{\lx@inpgf@ignorespaces\pi_{N}}f{\lx@inpgf@ignorespaces f}
Fig. 2: Commutative diagram of the key manifolds used to form a pullback bundle dynamical system. Note that the pullback bundle f∗​T​Nf^{*}TN is an nn-vector bundle over the robot configuration manifold MM.

III Pullback Bundle Dynamical Systems

Each task supplies a smooth map from robot configuration to a task manifold, a behavior metric, a potential, and dissipative forces. Pullback connections, metrics, and force operators express these task dynamics over robot configurations. Local auxiliary curves define desired task accelerations without requiring an invertible task map.

A weighted least-squares problem selects the robot acceleration that best matches all task accelerations. Nonzero task weights must be positive definite, and the combined active-task Jacobian must have full robot-coordinate rank. These assumptions make the local normal matrix invertible and yield a unique smooth policy. Separate weights can depend on velocity or deactivate tasks.

The stated stability theorem additionally requires strictly dissipative active forces and a proper combined potential. It gives compact invariant energy sublevels and convergence to an equilibrium set. This is not a guarantee that every initialization reaches the intended goal: undesired saddle or maximum equilibria can remain. The appendix explains the restricted-sublevel version when global properness is unavailable.

Task Task Map Behavior Metric Dissipative Forces Potential Weight Pseudometric
Goal Attraction f1:𝕊2→ℝf_{1}:\mathbb{S}^{2}\rightarrow\mathbb{R} : p↦dist​(p,pgoal)p\mapsto\textrm{dist}(p,p_{\textrm{goal}}) 𝒈1=1\bm{g}_{1}=1 𝓕D,1=0\bm{\mathcal{F}}_{D,1}=0 Φ1=∥𝒙∥22\Phi_{1}=\lVert\bm{x}\rVert^{2}_{2} 𝒘i=𝑰2\bm{w}_{i}=\bm{I}_{2}
Damping f2:𝕊2→𝕊2f_{2}:\mathbb{S}^{2}\rightarrow\mathbb{S}^{2} : p↦pp\mapsto p 𝒈¯2=𝑰3\bar{\bm{g}}_{2}=\bm{I}_{3} 𝓕¯D,2=−4​𝒙¯˙\bar{\bm{\mathcal{F}}}_{D,2}=-4\dot{\bar{\bm{x}}} Φ2=0\Phi_{2}=0 𝒘¯2=𝑰6\bar{\bm{w}}_{2}=\bm{I}_{6}
Obstacle Avoidance fi:𝕊2→ℝ+f_{i}:\mathbb{S}^{2}\rightarrow\mathbb{R}_{+} : p↦dist​(p,𝒳obs)p\mapsto\textrm{dist}(p,\mathcal{X}_{\textrm{obs}}) 𝒈i=exp⁡(1/(2​𝒙2))\bm{g}_{i}=\exp(1/(2\bm{x}^{2})) 𝓕D,i=0\bm{\mathcal{F}}_{D,i}=0 Φi=0\Phi_{i}=0 See (13)
TABLE I: Simple task design summary for running example of a robot on a sphere surface navigating to a goal point while avoiding obstacles. Note that each obstacle has an avoidance tasks (hence the indices ii). For the damping task, the overbars denote representations in ambient Euclidean space, which are easily pulled back through standard embedding maps to induce corresponding objects on 𝕊2\mathbb{S}^{2} and T​𝕊2T\mathbb{S}^{2} (see Appx. A.II for details).
Equations 2–9
f∗​T​N=∐p∈MπN−1​(f⁡(p)),f^{*}TN=\coprod_{p\in M}\pi_{N}^{-1}(f(p)), (2)
f∗​Γi​jk​(p)=∂fℓ∂xi​(p)​Γℓ​jk​(f⁡(p))f^{*}\Gamma^{k}_{ij}(p)=\frac{\partial f^{\ell}}{\partial x^{i}}(p)\Gamma^{k}_{\ell j}(f(p)) (3)
PBDSα(p,v)​{f∗​Dα(p,v)​γα(p,v)​(s)=f∗​ℱ​(γα(p,v)​(s))♯γα(p,v)​(0)=f∗​d​fα(p,v)​(0)​(α(p,v)′​(0)),α(p,v)′​(0)=(p,v).\textrm{PBDS}_{\alpha_{(p,v)}}\hskip-5.0pt\ \begin{cases}f^{*}D_{\alpha_{(p,v)}}\gamma_{\alpha_{(p,v)}}(s)=f^{*}\mathcal{F}(\gamma_{\alpha_{(p,v)}}(s))^{\sharp}\\ \gamma_{\alpha_{(p,v)}}(0)=f^{*}df_{\alpha_{(p,v)}(0)}(\alpha_{(p,v)}^{\prime}(0)),\\ \alpha_{(p,v)}^{\prime}(0)=(p,v).\end{cases}
S:T​M⟶T⁡(f∗​T​N):(p,v)↦πVB​(γα(p,v)′​(0)),S:TM\longrightarrow T(f^{*}TN):(p,v)\mapsto\pi_{\textrm{VB}}(\gamma^{\prime}_{\alpha_{(p,v)}}(0)), (4)
Zi:T​T​M\displaystyle Z_{i}:TTM ⟶T⁡(fi∗​T​Ni)\displaystyle\longrightarrow T(f_{i}^{*}TN_{i}) (5)
((p,v),a)\displaystyle\big((p,v),a\big) ↦πVB​(d​(fi∗​d​fi)(p,v)​(a)).\displaystyle\mapsto\pi_{\textrm{VB}}\Big(d(f_{i}^{*}df_{i})_{(p,v)}(a)\Big).
{σ¨​(t)=arg⁡mina∈𝒟σ˙​(t)​∑i=1K12​∥Zi​(a)−Si​(σ˙​(t))∥Fi∗​wi2σ˙​(0)=(p0,v0).\begin{cases}\ddot{\sigma}(t)=\underset{a\in\mathcal{D}_{\dot{\sigma}(t)}}{\arg\min}\ \sum^{K}_{i=1}\frac{1}{2}\lVert Z_{i}(a)-S_{i}(\dot{\sigma}(t))\rVert^{2}_{F_{i}^{*}w_{i}}\\ \dot{\sigma}(0)=(p_{0},v_{0}).\end{cases} (6)
𝝈¨=(∑i=1K𝑱​𝒇i⊤​𝒘ia​𝑱​𝒇i)†​(∑i=1K𝑱​𝒇i⊤​𝒘ia​𝓐i)\ddot{\bm{\sigma}}=\left(\sum_{i=1}^{K}\bm{Jf}_{i}^{\top}\bm{w}_{i}^{a}\bm{Jf}_{i}\right)^{\dagger}\left(\sum_{i=1}^{K}\bm{Jf}_{i}^{\top}\bm{w}_{i}^{a}\bm{\mathcal{A}}_{i}\right) (7)
𝓐i=𝒈i−1​(\displaystyle\bm{\mathcal{A}}_{i}=\bm{g}_{i}^{-1}( 𝓕D,i−∇𝚽i)−(𝑱​𝒇˙i−𝚵i)​𝝈˙\displaystyle\bm{\mathcal{F}}_{D,i}-\nabla\bm{\Phi}_{i})-(\dot{\bm{Jf}}_{i}-\bm{\Xi}_{i})\dot{\bm{\sigma}} (8)
(Ξi)k​j\displaystyle(\Xi_{i})_{kj} =(J​fi)ℓ​j​(Γi)ℓ​hk​(J​fi)h​r​σ˙r,\displaystyle=(Jf_{i})_{\ell j}(\Gamma_{i})^{k}_{\ell h}(Jf_{i})_{hr}\dot{\sigma}^{r},
V⁡(p,v)=∑i=1K12​∥(fi∗​d​fi)p​(v)∥(fi∗​gi)p2+Φi∘fi​(p).V(p,v)=\sum^{K}_{i=1}\frac{1}{2}\big\lVert(f_{i}^{*}df_{i})_{p}(v)\big\rVert^{2}_{(f_{i}^{*}g_{i})_{p}}+\Phi_{i}\circ f_{i}(p). (9)
M{\lx@inpgf@ignorespaces M}N1{\lx@inpgf@ignorespaces N_{1}}N2{\lx@inpgf@ignorespaces N_{2}}N3{\lx@inpgf@ignorespaces N_{3}}N1,1{\lx@inpgf@ignorespaces{N_{1,1}}}N1,2{\lx@inpgf@ignorespaces{N_{1,2}}}N1,3{\lx@inpgf@ignorespaces{N_{1,3}}}N3,1{\lx@inpgf@ignorespaces{N_{3,1}}}N3,2{\lx@inpgf@ignorespaces{N_{3,2}}}f2{\lx@inpgf@ignorespaces f_{2}}f3{\lx@inpgf@ignorespaces f_{3}}f1{\lx@inpgf@ignorespaces f_{1}}f1,1{\lx@inpgf@ignorespaces f_{1,1}}f1,2{\lx@inpgf@ignorespaces f_{1,2}}f1,3{\lx@inpgf@ignorespaces f_{1,3}}f3,1{\lx@inpgf@ignorespaces f_{3,1}}f3,2{\lx@inpgf@ignorespaces f_{3,2}}
Fig. 3: Tree of manifolds and task maps in a multi-task PBDS policy. At the root is the robot configuration manifold, and task manifolds are at the leaves. This structure can be exploited to parallelize computation of contributions from independent tasks and reuse computation from parent nodes.

IV Comparison to Riemannian Motion Policies on Geometric Dynamical Systems

The comparison examines coordinate invariance, the use of velocity-dependent metrics, and constraint design.

IV-A Geometric Consistency

The paper derives the compared GDS formulation from tangent-bundle dynamics and identifies a horizontal projection whose result depends on the chosen chart. Under a coordinate change, omitted metric terms generally reappear. Sphere experiments show differing GDS paths under alternative chart choices, while PBDS gives consistent physical trajectories. The claim concerns this specified formulation and construction.

Equation 10
𝑱​ϕ​(p^,v^)=[𝑱​ϕv​(p^,v^)0𝑱​ϕa​v​(p^,v^)𝑱​ϕa​(p^,v^)],J​ϕi​jv​(p^,v^)=J​ϕi​ja​(p^,v^)=∂p~j∂p^i​(p^),J​ϕi​ja​v​(p^,v^)=v^k​∂2p~j∂p^i​∂p^k​(p^).\begin{gathered}\bm{J\phi}(\hat{p},\hat{v})=\begin{bmatrix}\bm{J\phi}^{v}(\hat{p},\hat{v})&0\\ \bm{J\phi}^{av}(\hat{p},\hat{v})&\bm{J\phi}^{a}(\hat{p},\hat{v})\end{bmatrix},\\ J\phi^{v}_{ij}(\hat{p},\hat{v})=J\phi^{a}_{ij}(\hat{p},\hat{v})=\frac{\partial\tilde{p}^{j}}{\partial\hat{p}^{i}}(\hat{p}),\\ \vskip-4.0ptJ\phi^{av}_{ij}(\hat{p},\hat{v})=\hat{v}^{k}\frac{\partial^{2}\tilde{p}^{j}}{\partial\hat{p}^{i}\partial\hat{p}^{k}}(\hat{p}).\end{gathered} (10)
T​T​M{\lx@inpgf@ignorespaces TTM}F∗​T​T​N{\lx@inpgf@ignorespaces F^{*}TTN}T​T​N{\lx@inpgf@ignorespaces TTN}T​M{\lx@inpgf@ignorespaces TM}T​N{\lx@inpgf@ignorespaces TN}M{\lx@inpgf@ignorespaces M}N{\lx@inpgf@ignorespaces N}F∗​d​F{\lx@inpgf@ignorespaces F^{*}dF}πT​M{\lx@inpgf@ignorespaces\pi_{TM}}F∗​πT​N{\lx@inpgf@ignorespaces F^{*}\pi_{TN}}F∗​IdT​T​N{\lx@inpgf@ignorespaces F^{*}\textrm{Id}_{TTN}}πT​N{\lx@inpgf@ignorespaces\pi_{TN}}F{\lx@inpgf@ignorespaces F}πM{\lx@inpgf@ignorespaces\pi_{M}}πN{\lx@inpgf@ignorespaces\pi_{N}}f{\lx@inpgf@ignorespaces f}
Fig. 4: Commutative diagram for a PBDS system applied to a task map between tangent bundles.
Equation 11
𝒈^\displaystyle\hat{\bm{g}} =𝑱​ϕ⊤​𝒈~​𝑱​ϕ\displaystyle=\bm{J\phi}^{\top}\tilde{\bm{g}}\bm{J\phi} (11)
=[𝑱​ϕv⊤​𝒈~v​𝑱​ϕv+𝑱​ϕa​v⊤​𝒈~a​𝑱​ϕa​v𝑱​ϕa​v⊤​𝒈~a​𝑱​ϕa𝑱​ϕa⊤​𝒈~a​𝑱​ϕa​v𝑱​ϕa⊤​𝒈~a​𝑱​ϕa]\displaystyle=\begin{bmatrix}\bm{J\phi}^{v\top}\tilde{\bm{g}}^{v}\bm{J\phi}^{v}+\bm{J\phi}^{av\top}\tilde{\bm{g}}^{a}\bm{J\phi}^{av}&\bm{J\phi}^{av\top}\tilde{\bm{g}}^{a}\bm{J\phi}^{a}\\ \bm{J\phi}^{a\top}\tilde{\bm{g}}^{a}\bm{J\phi}^{av}&\bm{J\phi}^{a\top}\tilde{\bm{g}}^{a}\bm{J\phi}^{a}\end{bmatrix}

IV-B Task Prioritization and Velocity-Dependent Metrics

Using one velocity-dependent metric for both task behavior and priority can change effective inertia as well as force. Near a constraint, that can make acceleration away from the constraint difficult too. PBDS instead uses position-dependent behavior metrics and separate velocity-dependent weights. Appendix D derives a more elaborate tangent-bundle alternative, mainly to explain these design tradeoffs.

Equations
𝝈¨=(∑i𝓒i⊤​𝒘ia​𝓒i)†​(∑i𝓒i⊤​𝒘ia​𝓐i),𝓒i=𝑱​𝒇¯i+𝚵iv𝓐i=(𝒈i−1)a​(𝓕D,i−∇𝚽i)−(𝑱​𝒇¯˙i−𝚵ia)​𝝈˙Ξvi​j=JF^k​jΓi+dk⁡(h+m)Jf¯h​ℓσ˙ℓ,Ξai​j=JF^k⁡(j+d)Γi+dk⁡(h+m)Jf¯h​ℓσ˙ℓ.\begin{gathered}\textstyle\ddot{\bm{\sigma}}=\left(\sum_{i}\bm{\mathcal{C}}_{i}^{\top}\bm{w}_{i}^{a}\bm{\mathcal{C}}_{i}\right)^{\dagger}\left(\sum_{i}\bm{\mathcal{C}}_{i}^{\top}\bm{w}_{i}^{a}\bm{\mathcal{A}}_{i}\right),\quad\bm{\mathcal{C}}_{i}=\bm{J}\bar{\bm{f}}_{i}+\bm{\Xi}_{i}^{v}\vskip-2.0pt\\ \bm{\mathcal{A}}_{i}=(\bm{g}_{i}^{-1})^{a}(\bm{\mathcal{F}}_{D,i}-\nabla\bm{\Phi}_{i})-(\dot{\bm{J}\bar{\bm{f}}}_{i}-\bm{\Xi}_{i}^{a})\dot{\bm{\sigma}}\\ \Xi^{v}_{ij}=J\hat{F}_{kj}\Gamma^{i+d}_{k(h+m)}J\bar{f}_{h\ell}\dot{\sigma}^{\ell},\;\Xi^{a}_{ij}=J\hat{F}_{k(j+d)}\Gamma^{i+d}_{k(h+m)}J\bar{f}_{h\ell}\dot{\sigma}^{\ell}.\end{gathered}
σ¨=∑i(1+Ξia)​((gia)−1​(ℱD,i−∂xiΦi)−Ξiv​σ˙)∑i(1+Ξia)2Ξvi=12(gai)−1∂xigiaσ˙,Ξai=12(gai)−1∂vigiaσ˙.\begin{gathered}\ddot{\sigma}=\frac{\sum_{i}(1+\Xi_{i}^{a})((g^{a}_{i})^{-1}(\mathcal{F}_{D,i}-\partial_{x_{i}}\Phi_{i})-\Xi^{v}_{i}\dot{\sigma})}{\sum_{i}(1+\Xi^{a}_{i})^{2}}\\ \textstyle\Xi^{v}_{i}=\frac{1}{2}(g^{a}_{i})^{-1}\partial_{x_{i}}g_{i}^{a}\dot{\sigma},\quad\Xi^{a}_{i}=\frac{1}{2}(g^{a}_{i})^{-1}\partial_{v_{i}}g_{i}^{a}\dot{\sigma}.\end{gathered}

IV-C Metric-based Constraints

A constraint task maps configuration to positive distance from a boundary. An exponentially growing metric produces curvature that slows motion toward that boundary without adding a competing repulsive potential. A distance-and-velocity weight deactivates the task when it is irrelevant; smooth approximations are needed to retain the smooth-policy assumptions.

The prose specifies activation while approaching the boundary, where distance decreases. The displayed equation (13) prints the opposite velocity sign; the source equation is retained, so an implementation should reconcile that discrepancy. Avoiding spurious potential minima does not remove all possible equilibrium traps or certify arbitrary unsatisfiable task combinations.

Equations 12–13
σ¨=∑iJ​fi​wia​((gi)−1​(ℱD,i−∂xiΦi)−(J​f˙i+Ξi)​σ˙)∑iJ​fi2​wia.\ddot{\sigma}=\frac{\sum_{i}Jf_{i}w^{a}_{i}((g_{i})^{-1}(\mathcal{F}_{D,i}-\partial_{x_{i}}\Phi_{i})-(\dot{Jf}_{i}+\Xi_{i})\dot{\sigma})}{\sum_{i}Jf_{i}^{2}w^{a}_{i}}.
gi​(xi)=exp⁡(a/(b​xib))g_{i}(x_{i})=\exp(a/(bx_{i}^{b})) (12)
wia​(xi,x˙i)={1if ​x˙i>0​ and ​xi<β,0otherwise,w^{a}_{i}(x_{i},\dot{x}_{i})=\begin{cases}1&\mbox{if }\dot{x}_{i}>0\mbox{ and }x_{i}<\beta,\\ 0&\mbox{otherwise,}\end{cases} (13)
Refer to caption
(a) PBDS geometric consistency test
Refer to caption
(b) GDS/RMP geometric consistency test
Refer to caption
(c) PBDS attractor and obstacle avoidance policy on 𝕊2\mathbb{S}^{2}
Fig. 5: (a) and (b) show a geometric consistency test using a point attractor on 𝕊2\mathbb{S}^{2} and three schemes for choice of chart: (Red) Always using south pole stereographic projection, (Blue) always using north pole stereographic projection, (Purple) switching stereographic projections depending on the sphere hemisphere. The blue and green dots show the start and goal points, respectively. Note that using PBDS, all coordinate choices result in the same trajectory. (c) shows resulting trajectories from a PBDS attractor policy with obstacles added, where the curve colors represent different starting velocity directions.

V Implementation and Arm Experiments

The PBDS.jl implementation handles coordinate transitions and reuses task-map calculations in a tree. Algorithm 1 gives the policy evaluation steps. A seven-joint Franka Panda reaches through changing clutter to grasp a mug, combining attraction, obstacle and joint-limit tasks, and damping on sphere/circle task spaces. Policy outputs run at a reported 300–500 Hz; the experiment is a demonstration rather than a large statistical benchmark.

Algorithm 1 Multi-Task PBDS Policy
1: Input: Robot position and velocity (σ,σ˙)(\sigma,\dot{\sigma})
2: Output: Robot acceleration command σ¨\ddot{\sigma}
3: Data: Task PBDSs (fi,gi,Φi,ℱD,i)(f_{i},g_{i},\Phi_{i},\mathcal{F}_{D,i}), weights wiw_{i}
4: Compute 𝒜i\mathcal{A}_{i} using (8) for each task
5: Compute combined acceleration policy σ¨\ddot{\sigma} using (7)

VI Conclusions

PBDS provides a modular geometric language for reactive motion, separating task geometry from task priority. The stated results depend on rank, smoothness, dissipation, and potential assumptions. Velocity, acceleration, and actuator limits, as well as discrete mesh task spaces, are proposed extensions rather than general guarantees of the present method.

Appendix A Further Experimental Details and PBDS Design Strategies

The appendix documents the grasping task design. Joint limits make the Panda configuration model Euclidean; conservative sphere sets approximate link collision geometry.

I Task Set for Grasping a Mug

Tasks include joint damping; distance-toggled damping around the mug and along its rim; joint-limit, obstacle, and self-collision constraints; and approach/alignment/grasp attractors. A two-state machine delays grasping while the mug moves or the end effector is still approaching. Simultaneously active attractors are designed to be nonconflicting.

Refer to caption
Fig. 6: Example experimental scenario for Franka Panda arm reaching through clutter to grasp a mug. The blue and black boxes are obstacles, and the green and red spheres on the arm represent the collisions for different robot links.

II Tasks on 𝕊 n

Sphere tasks can be designed in their ambient Euclidean embedding and pulled back into north/south stereographic charts. The retained formulas give embeddings, transitions, and induced metrics, potentials, forces, and weights. This avoids treating one planar chart as globally valid on a sphere.

Equations 14–15
φ¯+i​(x)=φ¯−i​(x)=2​xia,a=∑i=1n1+(xi)2φ¯+n+1​(x)=−φ¯−n+1​(x)=2−aa.\begin{gathered}\bar{\varphi}_{+}^{i}(x)=\bar{\varphi}_{-}^{i}(x)=\frac{2x^{i}}{a},\ a=\sum_{i=1}^{n}1+(x^{i})^{2}\\ \bar{\varphi}_{+}^{n+1}(x)=-\bar{\varphi}_{-}^{n+1}(x)=\frac{2-a}{a}.\end{gathered} (14)
𝒈+=𝑱​𝝋¯+⊤​𝒈¯​𝑱​𝝋¯+,𝓕D+=𝑱​𝝋¯+⊤​𝓕¯D\displaystyle\bm{g}_{+}=\bm{J}\bar{\bm{\varphi}}_{+}^{\top}\bar{\bm{g}}\bm{J}\bar{\bm{\varphi}}_{+},\quad\bm{\mathcal{F}}_{D+}=\bm{J}\bar{\bm{\varphi}}_{+}^{\top}\bar{\bm{\mathcal{F}}}_{D} (15)
Φ+=Φ¯∘𝝋¯+,𝒘+a=𝑱φ¯+⊤𝒘¯a𝑱𝝋¯+.\displaystyle\Phi_{+}=\bar{\Phi}\bm{\circ}\bar{\bm{\varphi}}_{+},\quad\bm{w}_{+}^{a}=\bm{J}\bar{\varphi}_{+}^{\top}\bar{\bm{w}}^{a}\bm{J}\bar{\bm{\varphi}}_{+}.
ℝn+1{\lx@inpgf@ignorespaces\mathbb{R}^{n+1}}𝕊n{\lx@inpgf@ignorespaces\mathbb{S}^{n}}ℝn{\lx@inpgf@ignorespaces\mathbb{R}^{n}}ℝn{\lx@inpgf@ignorespaces\mathbb{R}^{n}}φ¯{\lx@inpgf@ignorespaces\bar{\varphi}}φ+{\lx@inpgf@ignorespaces\varphi_{+}}φ−{\lx@inpgf@ignorespaces\varphi_{-}}φ¯+{\lx@inpgf@ignorespaces\bar{\varphi}_{+}}φ−∘φ+−1{\lx@inpgf@ignorespaces\varphi_{-}\circ\varphi_{+}^{-1}}φ¯−{\lx@inpgf@ignorespaces\bar{\varphi}_{-}}
Fig. 7: Commutative diagram for the two stereographic charts of 𝕊2\mathbb{S}^{2} and its embedding in Euclidean space. (Here we omit the subsets U+,U−⊂𝕊2U_{+},U_{-}\subset\mathbb{S}^{2} which are the domains of the respective chart maps φ+,φ−\varphi_{+},\varphi_{-}).

III Distance-Toggled Tasks

Adding a distance coordinate to a task map lets its weight switch behavior on and off according to proximity. This is used both to skip unnecessary collision calculations and to activate appropriate damping or attraction near the mug.

IV A General Strategy for Multi-task PBDS Design

The suggested recipe starts with an identity-map damping task to cover every robot direction, then adds metric-only distance constraints, and finally adds nonconflicting attractors. This supports the rank and dissipation assumptions and avoids competing repulsive potentials. The authors explicitly distinguish absence of spurious potential minima from guaranteed goal completion at every initial state.

Appendix B Computational Tree for Multi-Task PBDS

Shared task-map prefixes make repeated Jacobian and derivative calculations wasteful. Leaf nodes compute weighted curvature, force, and Jacobian terms; intermediate nodes transform and combine them; the root solves for acceleration. The appendix expands a two-level tree to show that the recursive result matches the direct summed policy. All recurrence equations and the task-tree diagram are retained.

Equations 16–17
𝑷=𝑩=𝑱​𝒇⊤​𝒘a​𝑱​𝒇,𝑨=𝑱​𝒇⊤​𝒘a​𝑱​𝒇˙𝓕=𝑱​𝒇⊤​𝒘a​𝒈−1​(𝓕D−∇𝚽)ξs​rq=J​fτ​q​wτ​ηa​J​fα​s​Γα​βη​J​fβ​r,\begin{gathered}\bm{P}=\bm{B}=\bm{Jf}^{\top}\bm{w}^{a}\bm{Jf},\quad\bm{A}=\bm{Jf}^{\top}\bm{w}^{a}\dot{\bm{Jf}}\\ \bm{\mathcal{F}}=\bm{Jf}^{\top}\bm{w}^{a}\bm{g}^{-1}(\bm{\mathcal{F}}_{D}-\nabla\bm{\Phi})\\ \xi^{q}_{sr}=Jf_{\tau q}w^{a}_{\tau\eta}Jf_{\alpha s}\Gamma^{\eta}_{\alpha\beta}Jf_{\beta r},\end{gathered} (16)
𝑷=𝑱​𝒇⊤​(∑𝑷c)​𝑱​𝒇,𝑩=𝑱​𝒇⊤​(∑𝑩c)​𝑱​𝒇𝑨=𝑱​𝒇⊤​((∑𝑨c)​𝑱​𝒇+(∑𝑩c)​𝑱​𝒇˙)𝓕=𝑱𝒇⊤∑𝓕c,ξkh​ℓ=Jfq​kJfs​h(∑ξc)qs​rJfr​ℓ,\begin{gathered}\bm{P}=\bm{Jf}^{\top}\left(\sum\bm{P}_{c}\right)\bm{Jf},\quad\bm{B}=\bm{Jf}^{\top}\left(\sum\bm{B}_{c}\right)\bm{Jf}\\ \bm{A}=\bm{Jf}^{\top}\left(\left(\sum\bm{A}_{c}\right)\bm{Jf}+\left(\sum\bm{B}_{c}\right)\dot{\bm{Jf}}\right)\\ \bm{\mathcal{F}}=\bm{Jf}^{\top}\sum\bm{\mathcal{F}}_{c},\quad\xi^{k}_{h\ell}=Jf_{qk}Jf_{sh}\left(\sum\xi_{c}\right)^{q}_{sr}Jf_{r\ell},\end{gathered}\hskip-10.0pt (17)
M{\lx@inpgf@ignorespaces M}N1{\lx@inpgf@ignorespaces N_{1}}…{\lx@inpgf@ignorespaces\ldots}Nb{\lx@inpgf@ignorespaces N_{b}}N1,2{\lx@inpgf@ignorespaces{N_{1,2}}}…{\lx@inpgf@ignorespaces\ldots}N1,c1{\lx@inpgf@ignorespaces{N_{1,c_{1}}}}Nb,1{\lx@inpgf@ignorespaces{N_{b,1}}}…{\lx@inpgf@ignorespaces\ldots}Nb,cb{\lx@inpgf@ignorespaces{N_{b,c_{b}}}}fb{\lx@inpgf@ignorespaces f_{b}}f1{\lx@inpgf@ignorespaces f_{1}}f1,2{\lx@inpgf@ignorespaces f_{1,2}}f1,c1{\lx@inpgf@ignorespaces f_{1,c_{1}}}fb,1{\lx@inpgf@ignorespaces f_{b,1}}fb,cb{\lx@inpgf@ignorespaces f_{b,c_{b}}}
Fig. 8: Task map tree for an example multi-task PBDS.
Equations 18–19
𝑷=∑𝑷c,𝑨=∑𝑨c,𝓕=∑𝓕cξk​ℓ=(∑ξc)h​ℓk​σ˙h.\begin{gathered}\bm{P}=\sum\bm{P}_{c},\quad\bm{A}=\sum\bm{A}_{c},\quad\bm{\mathcal{F}}=\sum\bm{\mathcal{F}}_{c}\\ \xi_{k\ell}=\left(\sum\xi_{c}\right)^{k}_{h\ell}\dot{\sigma}^{h}.\end{gathered} (18)
𝝈¨=𝑷†​(𝓕−(𝑨+𝝃)​𝝈˙).\ddot{\bm{\sigma}}=\bm{P}^{\dagger}(\bm{\mathcal{F}}-(\bm{A}+\bm{\xi})\dot{\bm{\sigma}}). (19)
𝑷†​(\displaystyle\bm{P}^{\dagger}( 𝓕−(𝑨+𝝃)​𝝈˙)\displaystyle\bm{\mathcal{F}}-(\bm{A}+\bm{\xi})\dot{\bm{\sigma}})
=(∑i𝑱​𝒇i⊤​𝒘ia​𝑱​𝒇i)†​(∑i𝑱​𝒇i⊤​𝒘ia​𝓐i)\displaystyle=\left(\sum_{i}\bm{Jf}_{i}^{\top}\bm{w}_{i}^{a}\bm{Jf}_{i}\right)^{\dagger}\left(\sum_{i}\bm{Jf}_{i}^{\top}\bm{w}_{i}^{a}\bm{\mathcal{A}}_{i}\right)
𝝈¨=\displaystyle\ddot{\bm{\sigma}}= (∑i=1b∑j=1ci𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja​𝑱​𝒇i,j​𝑱​𝒇i)†\displaystyle\left(\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\bm{Jf}_{i,j}\bm{Jf}_{i}\right)^{\dagger}
(∑i=1b∑j=1ci𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja​(𝒈i,j−1​(𝓕D,i,j−∇Φi,j)\displaystyle\Bigg(\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\Bigg(\bm{g}_{i,j}^{-1}\Big(\bm{\mathcal{F}}_{D,i,j}-\nabla\Phi_{i,j}\Big)
−(dd​t​(𝑱​𝒇i,j​𝑱​𝒇i)+𝚵i,j)​𝝈˙)),\displaystyle-\Bigg(\frac{d}{dt}\Big(\bm{Jf}_{i,j}\bm{Jf}_{i}\Big)+\bm{\Xi}_{i,j}\Bigg)\dot{\bm{\sigma}}\Bigg)\Bigg),
𝑷\displaystyle\bm{P} =∑i=1b𝑷i=∑i=1b𝑱​𝒇i⊤​(∑j=1ci𝑷i,j)​𝑱​𝒇i\displaystyle=\sum_{i=1}^{b}\bm{P}_{i}=\sum_{i=1}^{b}\bm{Jf}_{i}^{\top}\bigg(\sum_{j=1}^{c_{i}}\bm{P}_{i,j}\bigg)\bm{Jf}_{i}
=∑i=1b∑j=1ci𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja​𝑱​𝒇i,j​𝑱​𝒇i,\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\bm{Jf}_{i,j}\bm{Jf}_{i},
𝑨=∑i=1b𝑨i\displaystyle\bm{A}=\sum_{i=1}^{b}\bm{A}_{i}
=∑i=1b𝑱​𝒇i⊤​((∑j=1ci𝑨i,j)​𝑱​𝒇i˙+(∑j=1ci𝑩i,j)​𝑱​𝒇i)\displaystyle=\sum_{i=1}^{b}\bm{Jf}_{i}^{\top}\Bigg(\bigg(\sum_{j=1}^{c_{i}}\bm{A}_{i,j}\bigg)\dot{\bm{Jf}_{i}}+\bigg(\sum_{j=1}^{c_{i}}\bm{B}_{i,j}\bigg)\bm{Jf}_{i}\Bigg)
=∑i=1b∑j=1ci𝑱​𝒇i⊤​(𝑱​𝒇i,j⊤​𝒘i,ja​𝑱​𝒇i,j​𝑱​𝒇i˙+𝑱​𝒇i⊤​𝒘i,ja​𝑱​𝒇i,j˙​𝑱​𝒇i)\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\bm{Jf}_{i}^{\top}\left(\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\bm{Jf}_{i,j}\dot{\bm{Jf}_{i}}+\bm{Jf}_{i}^{\top}\bm{w}_{i,j}^{a}\dot{\bm{Jf}_{i,j}}\bm{Jf}_{i}\right)
=∑i=1b∑j=1ci𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja​dd​t​(𝑱​𝒇i,j​𝑱​𝒇i),\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\frac{d}{dt}\Big(\bm{Jf}_{i,j}\bm{Jf}_{i}\Big),
𝓕\displaystyle\bm{\mathcal{F}} =∑i=1b𝓕i=∑i=1b𝑱​𝒇i⊤​(∑j=1ciℱi,j)\displaystyle=\sum_{i=1}^{b}\bm{\mathcal{F}}_{i}=\sum_{i=1}^{b}\bm{Jf}_{i}^{\top}\bigg(\sum_{j=1}^{c_{i}}\mathcal{F}_{i,j}\bigg)
=∑i=1b∑j=1ci𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja​𝒈i,j−1​(𝓕D,i,j−∇Φi,j),\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\bm{g}_{i,j}^{-1}\big(\bm{\mathcal{F}}_{D,i,j}-\nabla\Phi_{i,j}\big),
ξk​ℓ\displaystyle\xi_{k\ell} =(∑i=1bξi)h​ℓk​σ˙h\displaystyle=\bigg(\sum_{i=1}^{b}\xi_{i}\bigg)^{k}_{h\ell}\dot{\sigma}^{h}
=∑i=1b(J​fi)q​k​(J​fi)s​h​(∑j=1ciξi,j)s​rq​(J​fi)r​ℓ​σ˙h\displaystyle=\sum_{i=1}^{b}(Jf_{i})_{qk}(Jf_{i})_{sh}\bigg(\sum_{j=1}^{c_{i}}\xi_{i,j}\bigg)^{q}_{sr}(Jf_{i})_{r\ell}\dot{\sigma}^{h}
=∑i=1b∑j=1ci(J​fi)q​k​(J​fi,j)τ​q​(wi​ja)τ​η​(J​fi)s​h​(J​fi,j)α​s\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}(Jf_{i})_{qk}(Jf_{i,j})_{\tau q}(w^{a}_{ij})_{\tau\eta}(Jf_{i})_{sh}(Jf_{i,j})_{\alpha s}
(Γi,j)α​βη​(J​fi,j)β​r​(J​fi)r​ℓ​σ˙h\displaystyle\quad(\Gamma_{i,j})^{\eta}_{\alpha\beta}(Jf_{i,j})_{\beta r}(Jf_{i})_{r\ell}\dot{\sigma}^{h}
=∑i=1b∑j=1ci(𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja)k​η​(J​fi,j​J​fi)α​h\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\big(\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\big)_{k\eta}(Jf_{i,j}Jf_{i})_{\alpha h}
(Γi,j)α​βη​(J​fi,j​J​fi)β​ℓ​σ˙h\displaystyle\quad(\Gamma_{i,j})^{\eta}_{\alpha\beta}(Jf_{i,j}Jf_{i})_{\beta\ell}\dot{\sigma}^{h}
=∑i=1b∑j=1ci(𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja)k​η​J​(fi,j∘fi)α​h\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\big(\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\big)_{k\eta}J(f_{i,j}\circ f_{i})_{\alpha h}
(Γi,j)α​βη​J​(fi,j∘fi)β​ℓ​σ˙h\displaystyle\quad(\Gamma_{i,j})^{\eta}_{\alpha\beta}J(f_{i,j}\circ f_{i})_{\beta\ell}\dot{\sigma}^{h}
=∑i=1b∑j=1ci(𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja)k​η​(Ξi,j)η​h​σ˙h\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\big(\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\big)_{k\eta}(\Xi_{i,j})_{\eta h}\dot{\sigma}^{h}
=∑i=1b∑j=1ci𝑱​𝒇i⊤​𝑱​𝒇i,j⊤​𝒘i,ja​𝚵i,j​𝝈˙.\displaystyle=\sum_{i=1}^{b}\sum_{j=1}^{c_{i}}\bm{Jf}_{i}^{\top}\bm{Jf}_{i,j}^{\top}\bm{w}_{i,j}^{a}\bm{\Xi}_{i,j}\dot{\bm{\sigma}}.

Appendix C Mathematical Details for Pullback Bundle Dynamical Systems

This appendix constructs the pullback objects, establishes coordinate consistency and policy well-posedness, and presents the stated stability proof.

Equation 20
Γi​jk=12​gk​h​(∂igj​h+∂jgi​h−∂hgi​j).\Gamma^{k}_{ij}=\frac{1}{2}g^{kh}(\partial_{i}g_{jh}+\partial_{j}g_{ih}-\partial_{h}g_{ij}). (20)

I Constructions on the Pullback Bundle

Because a pullback bundle is not itself the task tangent bundle, the connection, acceleration, metric, force, and gradient operators are built explicitly. The connection coefficients are checked against chart-transition rules. Compatibility of the pulled-back metric and connection supplies the energy differentiation identity used later. The sharp/flat maps convert covectors and vectors, while dissipative forces retain their sign property under pullback.

Equations 21–35
f∗​d​f:T​M⟶f∗​T​N:(p,v)↦(p,π2​(d​fp​(v)))f^{*}df:TM\longrightarrow f^{*}TN:(p,v)\mapsto(p,\pi_{2}(df_{p}(v))) (21)
f∗​IdT​N:f∗​T​N⟶T​N:(p,v)↦(f⁡(p),v),f^{*}\textrm{Id}_{TN}:f^{*}TN\longrightarrow TN:(p,v)\mapsto(f(p),v), (22)
f∗​Γi​jk​(p)=∂fℓ∂xi​(p)​Γℓ​jk​(f⁡(p)).f^{*}\Gamma^{k}_{ij}(p)=\frac{\partial f^{\ell}}{\partial x^{i}}(p)\Gamma^{k}_{\ell j}(f(p)).
f∗∇:Γ(TM)×Γ(f∗TN)⟶Γ(f∗TN).f^{*}\nabla:\Gamma(TM)\times\Gamma(f^{*}TN)\longrightarrow\Gamma(f^{*}TN).
Γ~i​jk=Ap​k​∂yq∂y~i​(Aj​r​Γq​rp+∂Aj​p∂yq).\tilde{\Gamma}^{k}_{ij}=A^{pk}\frac{\partial y^{q}}{\partial\tilde{y}^{i}}\left(A_{jr}\Gamma^{p}_{qr}+\frac{\partial A_{jp}}{\partial y^{q}}\right). (23)
f∗​Γ~i​jk\displaystyle f^{*}\tilde{\Gamma}^{k}_{ij} =∂f~ℓ∂x~i​(p)​Γ~ℓ​jk​(f⁡(p))\displaystyle=\frac{\partial\tilde{f}^{\ell}}{\partial\tilde{x}^{i}}(p)\tilde{\Gamma}^{k}_{\ell j}(f(p))
=∂yq∂y~ℓ​(f⁡(p))​∂f~ℓ∂x~i​(p)​Ap​k​(f⁡(p))\displaystyle=\frac{\partial y^{q}}{\partial\tilde{y}^{\ell}}(f(p))\frac{\partial\tilde{f}^{\ell}}{\partial\tilde{x}^{i}}(p)A^{pk}(f(p))
(Aj​r​(f⁡(p))​Γq​rp​(f⁡(p))+∂Aj​p∂yq​(f⁡(p)))\displaystyle\quad\left(A_{jr}(f(p))\Gamma^{p}_{qr}(f(p))+\frac{\partial A_{jp}}{\partial y^{q}}(f(p))\right)
=∂xℓ∂x~i​(p)​∂fq∂xℓ​(p)​Bp​k​(p)\displaystyle=\frac{\partial x^{\ell}}{\partial\tilde{x}^{i}}(p)\frac{\partial f^{q}}{\partial x^{\ell}}(p)B^{pk}(p)
(Bj​r​(p)​Γq​rp​(f⁡(p))+∂Aj​p∂yq​(f⁡(p)))\displaystyle\quad\left(B_{jr}(p)\Gamma^{p}_{qr}(f(p))+\frac{\partial A_{jp}}{\partial y^{q}}(f(p))\right)
=∂xℓ∂x~i​(p)​Bp​k​(p)\displaystyle=\frac{\partial x^{\ell}}{\partial\tilde{x}^{i}}(p)B^{pk}(p)
(Bj​r​(p)​∂fq∂xℓ​(p)​Γq​rp​(f⁡(p))+∂Aj​p∂yq​(f⁡(p))​∂fq∂xℓ​(p))\displaystyle\quad\left(B_{jr}(p)\frac{\partial f^{q}}{\partial x^{\ell}}(p)\Gamma^{p}_{qr}(f(p))+\frac{\partial A_{jp}}{\partial y^{q}}(f(p))\frac{\partial f^{q}}{\partial x^{\ell}}(p)\right)
=∂xℓ∂x~i​(p)​Bp​k​(p)​(Bj​r​(p)​f∗​Γℓ​rp​(p)+∂Bj​p∂xℓ​(p)).\displaystyle=\frac{\partial x^{\ell}}{\partial\tilde{x}^{i}}(p)B^{pk}(p)\left(B_{jr}(p)f^{*}\Gamma^{p}_{\ell r}(p)+\frac{\partial B_{jp}}{\partial x^{\ell}}(p)\right).
f∗𝔛(σ)≜{V:I⟶f∗TN|\displaystyle f^{*}\mathfrak{X}(\sigma)\triangleq\{V:I\longrightarrow f^{*}TN\;|\; V​ is smooth,\displaystyle V\textrm{ is smooth,} (24)
V(t)∈f∗Eσ⁡(t),∀t∈I}.\displaystyle V(t)\in f^{*}E_{\sigma(t)},\forall t\in I\}.
f∗​Dσ:f∗​𝔛​(σ)⟶f∗​𝔛​(σ),f^{*}D_{\sigma}:f^{*}\mathfrak{X}(\sigma)\longrightarrow f^{*}\mathfrak{X}(\sigma),
f∗DσV(t)=(V˙k(t)+σ˙i(t)Vj(t)f∗Γi​jk(σ(t)))f∗∂k|σ⁡(t),f^{*}D_{\sigma}V(t)=\left(\dot{V}^{k}(t)+\dot{\sigma}^{i}(t)V^{j}(t)f^{*}\Gamma^{k}_{ij}(\sigma(t))\right)f^{*}\partial_{k}|_{\sigma(t)}, (25)
f∗​∂i(p)=(p,π2​(∂i(f⁡(p))).f^{*}\partial_{i}(p)=(p,\pi_{2}(\partial_{i}(f(p))). (26)
f∗​g:\displaystyle f^{*}g:\; f∗​T​N×f∗​T​N⟶ℝ\displaystyle f^{*}TN\times f^{*}TN\longrightarrow\mathbb{R}
((p,v1),(p,v2))↦gf⁡(p)​(v1,v2)\displaystyle((p,v_{1}),(p,v_{2}))\mapsto g_{f(p)}(v_{1},v_{2})
dd​t​gρ⁡(t)​(X⁡(t),X⁡(t))=2​gρ⁡(t)​(Dρ​X​(t),X⁡(t))\frac{d}{dt}g_{\rho(t)}(X(t),X(t))=2g_{\rho(t)}(D_{\rho}X(t),X(t)) (27)
dd​t​f∗\displaystyle\frac{d}{dt}f^{*} gσ⁡(t)​(V⁡(t),V⁡(t))\displaystyle g_{\sigma(t)}(V(t),V(t))\vskip-5.0pt
=dd​t​gf⁡(σ⁡(t))​(f∗​IdT​N​(V⁡(t)),f∗​IdT​N​(V⁡(t)))\displaystyle=\frac{d}{dt}g_{f(\sigma(t))}\left(f^{*}\textrm{Id}_{TN}(V(t)),f^{*}\textrm{Id}_{TN}(V(t))\right)
=dd​t​gρ⁡(t)​(W⁡(t),W⁡(t))\displaystyle=\frac{d}{dt}g_{\rho(t)}\left(W(t),W(t)\right)
=2​gρ⁡(t)​(Dρ​W​(t),W⁡(t)).\displaystyle=2g_{\rho(t)}\left(D_{\rho}W(t),W(t)\right). (28)
Dρ\displaystyle D_{\rho} W(t)=(W˙k(t)+ρ˙ℓ(t)Wj(t)Γℓ​jk(ρ(t)))∂k|ρ⁡(t)\displaystyle W(t)=\left(\dot{W}^{k}(t)+\dot{\rho}^{\ell}(t)W^{j}(t)\Gamma^{k}_{\ell j}(\rho(t))\right)\partial_{k}|_{\rho(t)}\vskip-5.0pt
=(W˙k(t)+∂fℓ∂xi(σ(t))σ˙i(t)Wj(t)Γℓ​jk(ρ(t)))∂k|ρ⁡(t)\displaystyle=\left(\dot{W}^{k}(t)+\frac{\partial f^{\ell}}{\partial x^{i}}(\sigma(t))\dot{\sigma}^{i}(t)W^{j}(t)\Gamma^{k}_{\ell j}(\rho(t))\right)\partial_{k}|_{\rho(t)}\vskip-5.0pt
=(W˙k(t)+σ˙i(t)Wj(t)f∗Γi​jk(σ(t)))∂k|ρ⁡(t)\displaystyle=\left(\dot{W}^{k}(t)+\dot{\sigma}^{i}(t)W^{j}(t)f^{*}\Gamma^{k}_{ij}(\sigma(t))\right)\partial_{k}|_{\rho(t)}\vskip-5.0pt
=f∗IdT​N((V˙k(t)+σ˙i(t)Vj(t)f∗Γi​jk(σ(t)))∂k|σ⁡(t))\displaystyle=f^{*}\textrm{Id}_{TN}\left(\left(\dot{V}^{k}(t)+\dot{\sigma}^{i}(t)V^{j}(t)f^{*}\Gamma^{k}_{ij}(\sigma(t))\right)\partial_{k}|_{\sigma(t)}\right)
=f∗​IdT​N​(f∗​Dσ​V​(t)).\displaystyle=f^{*}\textrm{Id}_{TN}(f^{*}D_{\sigma}V(t)).
dd​t​f∗​gσ⁡(t)​(V⁡(t),V⁡(t))\displaystyle\qquad\quad\frac{d}{dt}f^{*}g_{\sigma(t)}(V(t),V(t))
=2​gρ⁡(t)​(f∗​IdT​N​(f∗​Dσ​V​(t)),f∗​IdT​N​(V⁡(t)))\displaystyle=2g_{\rho(t)}\left(f^{*}\textrm{Id}_{TN}(f^{*}D_{\sigma}V(t)),f^{*}\textrm{Id}_{TN}(V(t))\right)
=2​f∗​gσ⁡(t)​(f∗​Dσ​V​(t),V⁡(t)).∎\displaystyle=2f^{*}g_{\sigma(t)}(f^{*}D_{\sigma}V(t),V(t)).\qed
f∗​T∗​N=∐p∈M(πN∗)−1​(f⁡(p)),f^{*}T^{*}N=\coprod_{p\in M}(\pi^{*}_{N})^{-1}(f(p)), (29)
♭:f∗​T​N⟶f∗​T∗​N:(p,v)↦f∗​gp​(v,⋅),\flat:f^{*}TN\longrightarrow f^{*}T^{*}N:(p,v)\mapsto f^{*}g_{p}(v,\cdot), (30)
♯:f∗​T∗​N⟶f∗​T​N:(p,ω)↦(p,ω)♭.\sharp:f^{*}T^{*}N\longrightarrow f^{*}TN:(p,\omega)\mapsto(p,\omega)^{\flat}. (31)
f∗​ℱD:f∗​T​N\displaystyle f^{*}\mathcal{F}_{D}:f^{*}TN ⟶f∗​T∗​N\displaystyle\longrightarrow f^{*}T^{*}N (32)
(p,v)\displaystyle(p,v) ↦(p,π2(ℱD(f(p),v)),\displaystyle\mapsto(p,\pi_{2}(\mathcal{F}_{D}(f(p),v)),
BΦ:M⟶f∗​T∗​N:p↦∂Φ∂yi​(f⁡(p))​f∗​d​yi|p,B_{\Phi}:M\longrightarrow f^{*}T^{*}N:p\mapsto\frac{\partial\Phi}{\partial y^{i}}(f(p))f^{*}dy^{i}|_{p},
BΦ​(p)\displaystyle B_{\Phi}(p) (f∗​d​fp​(v))=∂Φ∂yi​(f⁡(p))​f∗​d​yi|p​(f∗​d​fp​(v)))\displaystyle(f^{*}df_{p}(v))=\frac{\partial\Phi}{\partial y^{i}}(f(p))f^{*}dy^{i}|_{p}(f^{*}df_{p}(v)))
=∂Φ∂yi​(f⁡(p))​f∗​d​yi|p​(vj​∂fk∂xj​(p)​∂∂yk|f⁡(p))\displaystyle=\frac{\partial\Phi}{\partial y^{i}}(f(p))f^{*}dy^{i}|_{p}\left(v^{j}\frac{\partial f^{k}}{\partial x^{j}}(p)\left.\frac{\partial}{\partial y^{k}}\right|_{f(p)}\right)
=vj​∂fi∂xj​(p)​∂Φ∂yi​(f⁡(p))=vj​∂(Φ∘f)∂yi​(p)\displaystyle=v^{j}\frac{\partial f^{i}}{\partial x^{j}}(p)\frac{\partial\Phi}{\partial y^{i}}(f(p))=v^{j}\frac{\partial(\Phi\circ f)}{\partial y^{i}}(p)
=d​(Φ∘f)​(p,v),\displaystyle=d(\Phi\circ f)(p,v), (33)
f∗​grad​Φ:M⟶f∗​T​N:p↦BΦ​(p)♯.f^{*}\textrm{grad}\,\Phi:M\longrightarrow f^{*}TN:p\mapsto B_{\Phi}(p)^{\sharp}. (34)
f∗​ℱ​(⋅)♯:f∗​T​N\displaystyle f^{*}\mathcal{F}(\cdot)^{\sharp}:f^{*}TN ⟶f∗​T​N\displaystyle\longrightarrow f^{*}TN (35)
(p,v)\displaystyle(p,v) ↦f∗​ℱD​(p,v)♯−f∗​grad​Φ​(p).\displaystyle\mapsto f^{*}\mathcal{F}_{D}(p,v)^{\sharp}-f^{*}\textrm{grad}\,\Phi(p).

II Local Pullback Bundle Dynamical Systems

At each robot state and velocity, a short auxiliary curve defines local task-bundle dynamics. Such a curve exists in a local chart. Its derivative at the initial point depends only on that state and velocity, not on the arbitrary continuation of the curve, so the desired task-acceleration operator is well defined even with redundancy.

Equations 36
PBDSα(p,v)​{f∗​Dα(p,v)​γα(p,v)​(s)=f∗​ℱ​(γα(p,v)​(s))♯γα(p,v)​(0)=f∗​d​fα(p,v)​(0)​(α(p,v)′​(0)),α(p,v)′​(0)=(p,v).\textrm{PBDS}_{\alpha_{(p,v)}}\hskip-5.0pt\ \begin{cases}f^{*}D_{\alpha_{(p,v)}}\gamma_{\alpha_{(p,v)}}(s)=f^{*}\mathcal{F}(\gamma_{\alpha_{(p,v)}}(s))^{\sharp}\\ \gamma_{\alpha_{(p,v)}}(0)=f^{*}df_{\alpha_{(p,v)}(0)}(\alpha_{(p,v)}^{\prime}(0)),\\ \alpha_{(p,v)}^{\prime}(0)=(p,v).\end{cases}
G:T​M⟶T⁡(f∗​T​N):(p,v)↦γ˙α(p,v)​(0).G:TM\longrightarrow T(f^{*}TN):(p,v)\mapsto\dot{\gamma}_{\alpha_{(p,v)}}(0). (36)
γ˙α(p,v)k​(0)=\displaystyle\dot{\gamma}_{\alpha_{(p,v)}}^{k}(0)=\; (f∗​ℱ​(γα(p,v)​(0))♯)k\displaystyle(f^{*}\mathcal{F}(\gamma_{\alpha_{(p,v)}}(0))^{\sharp})^{k}
−α˙(p,v)i​(0)​γα(p,v)j​(0)​f∗​Γi​jk​(α(p,v)​(0))\displaystyle-\dot{\alpha}_{(p,v)}^{i}(0)\gamma_{\alpha_{(p,v)}}^{j}(0)f^{*}\Gamma^{k}_{ij}(\alpha_{(p,v)}(0))
=\displaystyle=\; (f∗​ℱ​(f∗​d​f​(α˙(p,v)​(0)))♯)k\displaystyle(f^{*}\mathcal{F}(f^{*}df(\dot{\alpha}_{(p,v)}(0)))^{\sharp})^{k}
−α˙(p,v)i​(0)​(f∗​d​f​(α˙(p,v)​(0)))j​f∗​Γi​jk​(α(p,v)​(0))\displaystyle-\dot{\alpha}_{(p,v)}^{i}(0)(f^{*}df(\dot{\alpha}_{(p,v)}(0)))^{j}f^{*}\Gamma^{k}_{ij}(\alpha_{(p,v)}(0))
=\displaystyle=\; (f∗​ℱ​(γβ(p,v)​(0))♯)k\displaystyle(f^{*}\mathcal{F}(\gamma_{\beta_{(p,v)}}(0))^{\sharp})^{k}
−β˙(p,v)i​(0)​γβ(p,v)j​(0)​f∗​Γi​jk​(β(p,v)​(0))\displaystyle-\dot{\beta}_{(p,v)}^{i}(0)\gamma_{\beta_{(p,v)}}^{j}(0)f^{*}\Gamma^{k}_{ij}(\beta_{(p,v)}(0))
=\displaystyle=\; γ˙β(p,v)k​(0).\displaystyle\dot{\gamma}_{\beta_{(p,v)}}^{k}(0).

III Multi-Task Pullback Bundle Dynamical Systems

Vertical projections map robot and desired task accelerations into comparable spaces. The resulting optimization is a weighted quadratic least-squares problem. Positive active weights and a full-rank combined Jacobian make its normal matrix positive definite. A transversality argument discusses generic rank when sufficiently many tasks are available; practical designs still need the rank condition to hold.

Equations 37–46
Si:T​M\displaystyle S_{i}:TM ⟶T⁡(fi∗​T​Ni)\displaystyle\longrightarrow T\big(f_{i}^{*}TN_{i}\big) (37)
(p,v)\displaystyle(p,v) ↦πVB​(Gi​(p,v))=πVB​(γ˙vp,i​(0))\displaystyle\mapsto\pi_{\textrm{VB}}(G_{i}(p,v))=\pi_{\textrm{VB}}(\dot{\gamma}_{v_{p},i}(0))
=((fi∗​d​fi)p​(v),(0,γ˙vp,ia​(0))),\displaystyle\hskip 73.0pt=((f_{i}^{*}df_{i})_{p}(v),(0,\dot{\gamma}^{a}_{{v_{p}},i}(0))),
Zi:T​T​M\displaystyle Z_{i}:TTM ⟶T⁡(fi∗​T​Ni)\displaystyle\longrightarrow T(f_{i}^{*}TN_{i}) (38)
((p,v),a)\displaystyle\big((p,v),a\big) ↦πVB​(d​(fi∗​d​fi)(p,v)​(a)).\displaystyle\mapsto\pi_{\textrm{VB}}\Big(d(f_{i}^{*}df_{i})_{(p,v)}(a)\Big).
Fi:T​M\displaystyle F_{i}:TM ⟶T​Ni\displaystyle\longrightarrow TN_{i} (39)
(p,v)\displaystyle(p,v) ↦(fi​(p),(d​fi)p​(v)),\displaystyle\mapsto(f_{i}(p),(df_{i})_{p}(v)),
Fi∗​wi:\displaystyle F_{i}^{*}w_{i}: T⁡(fi∗​T​Ni)×T⁡(fi∗​T​Ni)⟶ℝ\displaystyle T(f_{i}^{*}TN_{i})\times T(f_{i}^{*}TN_{i})\longrightarrow\mathbb{R} (40)
((p,v),a1),((p,v),a2)↦(wi)Fi​(p,v)​(a1,a2).\displaystyle((p,v),a_{1}),((p,v),a_{2})\mapsto(w_{i})_{F_{i}(p,v)}(a_{1},a_{2}).
ζ:T​M\displaystyle\zeta:TM ⟶𝒟⊂T​T​M\displaystyle\longrightarrow\mathcal{D}\subset TTM (41)
(p,v)\displaystyle(p,v) ↦arg⁡mina∈𝒟(p,v)​∑i=1K12​∥Zi​(a)−Si​(p,v)∥Fi∗​wi2,\displaystyle\mapsto\underset{a\in\mathcal{D}_{(p,v)}}{\arg\min}\ \sum^{K}_{i=1}\frac{1}{2}\lVert Z_{i}(a)-S_{i}(p,v)\rVert^{2}_{F_{i}^{*}w_{i}},
𝑱⁡(𝒇i∗​𝒅​𝒇i)​(p,v)=[𝑱​𝒇i​(p)0𝑱​𝒇˙i​(p,v)𝑱​𝒇i​(p)],\bm{J}(\bm{f}_{i}^{*}\bm{df}_{i})(p,v)=\begin{bmatrix}\bm{Jf}_{i}(p)&0\\ \dot{\bm{Jf}}_{i}(p,v)&\bm{Jf}_{i}(p)\end{bmatrix}, (42)
(J​fi˙)k​j​(p,v)=vℓ​∂2fij∂xℓ​∂xk​(p).(\dot{Jf_{i}})_{kj}(p,v)=v^{\ell}\frac{\partial^{2}f^{j}_{i}}{\partial x^{\ell}\partial x^{k}}(p).
Zi\displaystyle Z_{i} (a)=πVB​(d​(fi∗​d​fi)(p,v)​(a))\displaystyle(a)=\pi_{\textrm{VB}}\left(d(f_{i}^{*}df_{i})_{(p,v)}(a)\right)
=πVB(((fi∗dfi)p(v),((Jfi)j​k(p)vk,\displaystyle=\pi_{\textrm{VB}}\big(((f_{i}^{*}df_{i})_{p}(v),((Jf_{i})_{jk}(p)v^{k},
(J​f˙i)j​k(p,v)vk+(Jfi)j​k(p)(aa)k)))\displaystyle\qquad\qquad(\dot{Jf}_{i})_{jk}(p,v)v^{k}+(Jf_{i})_{jk}(p)(a^{a})^{k}))\big)
=((fi∗​d​fi)p​(v),(0,(J​f˙i)j​k​(p,v)​vk+(J​fi)j​k​(p)​(aa)k).\displaystyle=((f_{i}^{*}df_{i})_{p}(v),(0,(\dot{Jf}_{i})_{jk}(p,v)v^{k}+(Jf_{i})_{jk}(p)(a^{a})^{k}).
∥Zi​(a)−Si​(p,v)∥Fi∗​wi2\displaystyle\lVert Z_{i}(a)-S_{i}(p,v)\rVert^{2}_{F_{i}^{*}w_{i}} (43)
=∥𝑱​𝒇i​(p)​𝒂a+𝑱​𝒇˙i​(p,v)​𝒗−𝜸˙vp,ia​(0)∥𝒘ia​(Fi​(p,v))2.\displaystyle=\lVert\bm{Jf}_{i}(p)\bm{a}^{a}+\dot{\bm{Jf}}_{i}(p,v)\bm{v}-\dot{\bm{\gamma}}^{a}_{v_{p},i}(0)\rVert^{2}_{\bm{w}_{i}^{a}(F_{i}(p,v))}.
(γ˙vp,ia)k\displaystyle(\dot{\gamma}^{a}_{v_{p},i})^{k} (0)=(fi∗​ℱi​(γvp,i​(0))♯)k−vℓ​(t)​γvp,ij​(0)​fi∗​(Γi)ℓ​jk​(p)\displaystyle(0)=(f_{i}^{*}\mathcal{F}_{i}(\gamma_{v_{p},i}(0))^{\sharp})^{k}-v^{\ell}(t)\gamma^{j}_{v_{p},i}(0)f_{i}^{*}(\Gamma_{i})^{k}_{\ell j}(p)
=fi∗​ℱD,i​(γvp,i​(0))♯−fi∗​grad​Φi​(p)\displaystyle=f^{*}_{i}\mathcal{F}_{D,i}(\gamma_{v_{p},i}(0))^{\sharp}-f^{*}_{i}\textrm{grad}\,\Phi_{i}(p)
−vℓ​(t)​γvp,ij​(0)​fi∗​(Γi)ℓ​jk​(p)\displaystyle\quad-v^{\ell}(t)\gamma^{j}_{v_{p},i}(0)f_{i}^{*}(\Gamma_{i})^{k}_{\ell j}(p)
=gik​j​(fi​(p))​(ℱD,ij​((d​fi)p​(v))−∂Φi∂yij​(fi​(p)))\displaystyle=g_{i}^{kj}(f_{i}(p))\bigg(\mathcal{F}_{D,i}^{j}((df_{i})_{p}(v))-\frac{\partial\Phi_{i}}{\partial y^{j}_{i}}(f_{i}(p))\bigg)
−vℓ​(t)​(J​fi)j​h​(p)​vh​(t)​(J​fi)r​ℓ​(p)​(Γi)r​jk​(fi​(p)).\displaystyle\quad-v^{\ell}(t)(Jf_{i})_{jh}(p)v^{h}(t)(Jf_{i})_{r\ell}(p)(\Gamma_{i})_{rj}^{k}(f_{i}(p)).
𝜻(p,v)=(∑i=1K\displaystyle\bm{\zeta}(p,v)=\Bigg(\sum_{i=1}^{K} 𝑱​𝒇i​(p)⊤​𝒘ia​(Fi​(p,v))​𝑱​𝒇i​(p))†\displaystyle\bm{Jf}_{i}(p)^{\top}\bm{w}_{i}^{a}(F_{i}(p,v))\bm{Jf}_{i}(p)\Bigg)^{\dagger} (44)
(∑i=1K𝑱​𝒇i​(p)⊤​𝒘ia​(Fi​(p,v))​𝓐i​(p,v)),\displaystyle\left(\sum_{i=1}^{K}\bm{Jf}_{i}(p)^{\top}\bm{w}_{i}^{a}(F_{i}(p,v))\bm{\mathcal{A}}_{i}(p,v)\right),
ℐ:T​M\displaystyle\mathcal{I}:TM ⟶𝒫⁡({1,…,K})\displaystyle\longrightarrow\mathcal{P}(\{1,\ldots,K\}) (45)
(p,v)\displaystyle(p,v) ↦{i∈{1,…,K}|wi​((d​fi)p​(v))≠0}.\displaystyle\mapsto\{i\in\mathbb{\{}1,\ldots,K\}\ |\ w_{i}((df_{i})_{p}(v))\neq 0\}.
∑i=1K𝑱​𝒇i​(p)⊤​𝒘ia​(Fi​(p,v))​𝑱​𝒇i​(p)\displaystyle\qquad\sum_{i=1}^{K}\bm{Jf}_{i}(p)^{\top}\bm{w}_{i}^{a}(F_{i}(p,v))\bm{Jf}_{i}(p)
=∑i∈ℐ⁡(p,v)𝑱​𝒇i​(p)⊤​𝒘ia​(Fi​(p,v))​𝑱​𝒇i​(p)\displaystyle=\sum_{i\in\mathcal{I}(p,v)}\bm{Jf}_{i}(p)^{\top}\bm{w}_{i}^{a}(F_{i}(p,v))\bm{Jf}_{i}(p)
=𝑱​𝒇(p,v)⊤​(p)​𝒘(p,v)a​(F(p,v)​(p,v))​𝑱​𝒇(p,v)​(p),\displaystyle=\bm{Jf}_{(p,v)}^{\top}(p)\bm{w}^{a}_{(p,v)}(F_{(p,v)}(p,v))\bm{Jf}_{(p,v)}(p),
𝒮≜{\displaystyle\mathcal{S}\triangleq\Big\{ (f1,…,f2​m)∈C∞(M;N1)×\displaystyle(f_{1},\dots,f_{2m})\in C^{\infty}(M;N_{1})\times
⋯×C∞(M;N2​m):rank(f2​m)<m}\displaystyle\dots\times C^{\infty}(M;N_{2m}):\ \textrm{rank}(f^{2m})<m\Big\}
θ⁡(f1,…,f2​m)≜(θ~,θ0f1,…,θ0f2​m,θ1f1,…,θ1f2​m)\theta(f_{1},\dots,f_{2m})\triangleq\Big(\tilde{\theta},\theta^{f_{1}}_{0},\dots,\theta^{f_{2m}}_{0},\theta^{f_{1}}_{1},\dots,\theta^{f_{2m}}_{1}\Big)
ρ⁡(θ)≜(θ1f1,…,θ1f2​m).\rho(\theta)\triangleq\Big(\theta^{f_{1}}_{1},\dots,\theta^{f_{2m}}_{1}\Big).
𝒮=⋃1≤i≤m−1​ρ−1​(L⁡(2​m,m,i)),\mathcal{S}=\underset{1\leq i\leq m-1}{\bigcup}\rho^{-1}(L(2m,m;i)),
codim​(𝒮)\displaystyle\textrm{codim}(\mathcal{S}) ≥codim​(ρ−1​(L⁡(2​m,m,m−1)))\displaystyle\geq\textrm{codim}\big(\rho^{-1}(L(2m,m;m-1))\big)
=codim​(L⁡(2​m,m,m−1))=m+1.\displaystyle=\textrm{codim}\big(L(2m,m;m-1)\big)=m+1.
{σ¨​(t)=ζ​(σ˙​(t))=arg⁡mina∈𝒟σ˙​(t)​∑i=1K12​∥Zi​(a)−Si​(σ˙​(t))∥Fi∗​wi2σ˙​(0)=(p0,v0).\begin{cases}\ddot{\sigma}(t)=\zeta(\dot{\sigma}(t))=\\ \qquad\underset{a\in\mathcal{D}_{\dot{\sigma}(t)}}{\arg\min}\ \sum^{K}_{i=1}\frac{1}{2}\lVert Z_{i}(a)-S_{i}(\dot{\sigma}(t))\rVert^{2}_{F_{i}^{*}w_{i}}\\ \dot{\sigma}(0)=(p_{0},v_{0}).\end{cases}\hskip-10.0pt (46)

IV Proof of Multi-Task PBDS Stability

The paper sums task kinetic energies and potentials, then uses metric–connection compatibility to cancel conservative terms. Strict dissipation yields decreasing energy away from zero velocity. Properness gives compact invariant sublevels and forward completeness, and LaSalle’s principle identifies the stated limiting equilibrium set.

If properness holds only on a restricted sublevel, so does the conclusion. Equilibria at attractor-potential maxima or saddles may remain, even without unwanted local minima; the authors suggest small perturbations to escape such cases. The preserved derivation records the exact conditions and claim.

Equations 47–49
V⁡(p,v)=∑i=1K12​∥(fi∗​d​fi)p​(v)∥(fi∗​gi)p2+Φi∘fi​(p),V(p,v)=\sum^{K}_{i=1}\frac{1}{2}\big\lVert(f_{i}^{*}df_{i})_{p}(v)\big\rVert^{2}_{(f_{i}^{*}g_{i})_{p}}+\Phi_{i}\circ f_{i}(p), (47)
d​Vd​t​(σ˙​(t))=∑i=1K⟨ℱD,i​((d​fi)σ⁡(t)​(σ˙​(t))),(d​fi)σ⁡(t)​(σ˙​(t))⟩<0\frac{dV}{dt}(\dot{\sigma}(t))=\sum^{K}_{i=1}\langle\mathcal{F}_{D,i}((df_{i})_{\sigma(t)}(\dot{\sigma}(t))),(df_{i})_{\sigma(t)}(\dot{\sigma}(t))\rangle<0
dd​t​(12​∥(fi∗​d​fi)σ⁡(t)​(σ˙​(t))∥(fi∗​gi)σ⁡(t)2)\displaystyle\frac{d}{dt}\Big(\frac{1}{2}\big\lVert(f_{i}^{*}df_{i})_{\sigma(t)}(\dot{\sigma}(t))\big\rVert^{2}_{(f_{i}^{*}g_{i})_{\sigma(t)}}\Big)
=dd​t​(12​(fi∗​gi)σ⁡(t)​((fi∗​d​fi)σ⁡(t)​(σ˙​(t)),(fi∗​d​fi)σ⁡(t)​(σ˙​(t))))\displaystyle=\frac{d}{dt}\Big(\frac{1}{2}(f_{i}^{*}g_{i})_{\sigma(t)}\big((f_{i}^{*}df_{i})_{\sigma(t)}({\dot{\sigma}(t)}),(f_{i}^{*}df_{i})_{\sigma(t)}(\dot{\sigma}(t))\big)\Big)
=(fi∗​gi)σ⁡(t)​(fi∗​Dσ,i​γσ,i​(0),γσ,i​(0))\displaystyle=(f_{i}^{*}g_{i})_{\sigma(t)}(f_{i}^{*}D_{\sigma,i}\gamma_{\sigma,i}(0),\gamma_{\sigma,i}(0)) (48)
=(fi∗​gi)σ⁡(t)​(fi∗​ℱD,i​(γσ,i​(0))♯,γσ,i​(0))\displaystyle=(f_{i}^{*}g_{i})_{\sigma(t)}(f_{i}^{*}\mathcal{F}_{D,i}(\gamma_{\sigma,i}(0))^{\sharp},\gamma_{\sigma,i}(0))
−(fi∗​gi)σ⁡(t)​(f∗​grad​Φi​(σ⁡(t)),γσ,i​(0))\displaystyle\quad-(f_{i}^{*}g_{i})_{\sigma(t)}(f^{*}\textrm{grad}\,\Phi_{i}(\sigma(t)),\gamma_{\sigma,i}(0))
=⟨fi∗​ℱD,i​(γσ,i​(0)),γσ,i​(0)⟩\displaystyle=\langle f_{i}^{*}\mathcal{F}_{D,i}(\gamma_{\sigma,i}(0)),\gamma_{\sigma,i}(0)\rangle
−⟨fi∗​grad​Φi​(σ⁡(t))♭,γσ,i​(0)⟩\displaystyle\quad-\langle f_{i}^{*}\textrm{grad}\Phi_{i}(\sigma(t))^{\flat},\gamma_{\sigma,i}(0)\rangle
=⟨ℱD,i​((d​fi)σ⁡(t)​(σ˙​(t))),(d​fi)σ⁡(t)​(σ˙​(t))⟩\displaystyle=\langle\mathcal{F}_{D,i}((df_{i})_{\sigma(t)}(\dot{\sigma}(t))),(df_{i})_{\sigma(t)}(\dot{\sigma}(t))\rangle
−BΦi​(σ⁡(t))​((fi∗​d​fi)σ⁡(t)​(σ˙​(t)))\displaystyle\quad-B_{\Phi_{i}}(\sigma(t))\big((f_{i}^{*}df_{i})_{\sigma(t)}(\dot{\sigma}(t))\big)
=⟨ℱD,i​((d​fi)σ⁡(t)​(σ˙​(t))),(d​fi)σ⁡(t)​(σ˙​(t))⟩\displaystyle=\langle\mathcal{F}_{D,i}((df_{i})_{\sigma(t)}(\dot{\sigma}(t))),(df_{i})_{\sigma(t)}(\dot{\sigma}(t))\rangle (49)
−d⁡(Φi∘fi)​(σ˙​(t)),\displaystyle\quad-d(\Phi_{i}\circ f_{i})(\dot{\sigma}(t)),
dd​t​(Φi∘fi​(σ⁡(t)))=d⁡(Φi∘fi)​(σ˙​(t)).\frac{d}{dt}\Big(\Phi_{i}\circ f_{i}(\sigma(t))\Big)=d(\Phi_{i}\circ f_{i})(\dot{\sigma}(t)).
dd​t​(V⁡(σ˙​(t)))=∑i=1K⟨ℱD,i​((d​fi)σ⁡(t)​(σ˙​(t))),(d​fi)σ⁡(t)​(σ˙​(t))⟩.\frac{d}{dt}\big(V(\dot{\sigma}(t))\big)=\sum^{K}_{i=1}\langle\mathcal{F}_{D,i}((df_{i})_{\sigma(t)}(\dot{\sigma}(t))),(df_{i})_{\sigma(t)}(\dot{\sigma}(t))\rangle.
f∗​g:T​M×T​M\displaystyle f^{*}g:TM\times TM ⟶ℝ\displaystyle\longrightarrow\mathbb{R}
((p,v1),(p,v2))\displaystyle((p,v_{1}),(p,v_{2})) ↦∑i=1K(fi∗​gi)p​((fi∗​d​fi)p​(v1),(fi∗​d​fi)p​(v2)).\displaystyle\mapsto\sum^{K}_{i=1}(f_{i}^{*}g_{i})_{p}\big((f_{i}^{*}df_{i})_{p}(v_{1}),(f_{i}^{*}df_{i})_{p}(v_{2})\big).
A\displaystyle A ={(p,v)∈V−1([0,β]):\displaystyle=\Big\{(p,v)\in V^{-1}([0,\beta]):
⟨ℱD,(p,v)((df(p,v))p(v)),(df(p,v))p(v)⟩=0}\displaystyle\hskip 30.0pt\langle\mathcal{F}_{D,(p,v)}\big((df_{(p,v)})_{p}(v)\big),(df_{(p,v)})_{p}(v)\rangle=0\Big\}
={(p,0)∈V−1([0,β])}.\displaystyle=\Big\{(p,0)\in V^{-1}([0,\beta])\Big\}.
B={(p,0)∈\displaystyle B=\Big\{(p,0)\in V−1​([0,β]):\displaystyle V^{-1}([0,\beta]):
fi∗gradΦi(p)=0,wi(fi(p),0)≠0}\displaystyle f_{i}^{*}\textrm{grad}\Phi_{i}(p)=0,\ w_{i}(f_{i}(p),0)\neq 0\Big\}

Appendix D Mathematical Details for PBDS Policies on Tangent Bundle Tasks

Extending PBDS from configuration tasks to tangent-bundle tasks introduces extra velocity coordinates and jerk terms. The appendix explores a constrained construction that still produces acceleration commands. Its coordinate-swapping step requires appropriate parallelizable embedded manifolds, a restriction that excludes the sphere example in this form.

I Enforcing Velocity Constraints

An admissible curve in the tangent bundle must have its position derivative equal its represented velocity. An augmented task map and a subbundle constraint enforce that relation. A permutation enables a globally defined vertical projection to remove the jerk components, and a modified connection keeps evolution within the required subbundle.

Equations 50
F×πM:T​M\displaystyle F\times\pi_{M}:TM ⟶T​N×M\displaystyle\longrightarrow TN\times M
(p,v)\displaystyle(p,v) ↦(F⁡(p,v),πM​(p,v))=((p,d​fp​(v)),p).\displaystyle\mapsto(F(p,v),\pi_{M}(p,v))=((p,df_{p}(v)),p).
F^≜σT​N∘F:T​M\displaystyle\hat{F}\triangleq\sigma_{TN}\circ F:TM ⟶ℝ2​d\displaystyle\longrightarrow\mathbb{R}^{2d}
(p,v)\displaystyle(p,v) ↦(d​f¯p​(v),f¯​(p))\displaystyle\mapsto(d\bar{f}_{p}(v),\bar{f}(p))
F˘≜F^×πM:T​M\displaystyle\breve{F}\triangleq\hat{F}\times\pi_{M}:TM ⟶ℝ2​d×M\displaystyle\longrightarrow\mathbb{R}^{2d}\times M
(p,v)\displaystyle(p,v) ↦(d​f¯p​(v),f¯​(p),p).\displaystyle\mapsto(d\bar{f}_{p}(v),\bar{f}(p),p).
F^∗​ℒ:T​T​M\displaystyle\hat{F}^{*}\mathcal{L}:TTM ⟶F^∗​T​ℝ2​d⊕πM∗​T​M\displaystyle\longrightarrow\hat{F}^{*}T\mathbb{R}^{2d}\oplus\pi_{M}^{*}TM
((p,v),(b,a))\displaystyle((p,v),(b,a)) ↦((p,v),\displaystyle\mapsto((p,v),
d​F^(p,v)​(b,a)+((d​πM)(p,v)​(b,a)−v)).\displaystyle\qquad d\hat{F}_{(p,v)}(b,a)+((d\pi_{M})_{(p,v)}(b,a)-v)).
F˘∗GVB={((p,v),ci1∂xℝ2​di+ci2∂vℝ2​di+ci3∂xMi)\displaystyle\breve{F}^{*}G_{\textrm{VB}}=\Big\{((p,v),c^{1}_{i}\partial x^{i}_{\mathbb{R}^{2d}}+c^{2}_{i}\partial v^{i}_{\mathbb{R}^{2d}}+c_{i}^{3}\partial x^{i}_{M})
∈F^∗​T​ℝ2​d⊕πM∗​T​M\displaystyle\qquad\quad\in\hat{F}^{*}T\mathbb{R}^{2d}\oplus\pi_{M}^{*}TM
|ci1=0 for i=1,…,d,ci3=0 for i=1,…,m},\displaystyle\qquad\quad|\ c_{i}^{1}=0\textrm{ for }i=1,\ldots,d,c_{i}^{3}=0\textrm{ for }i=1,\ldots,m\Big\},
F˘∗Gℝ2​d={((p,v),ci1∂xℝ2​di+ci2∂vℝ2​di+ci3∂xMi)\displaystyle\breve{F}^{*}G_{\mathbb{R}^{2d}}=\Big\{((p,v),c^{1}_{i}\partial x^{i}_{\mathbb{R}^{2d}}+c^{2}_{i}\partial v^{i}_{\mathbb{R}^{2d}}+c_{i}^{3}\partial x^{i}_{M})
∈F^∗Tℝ2​d⊕πM∗TM|ci3=0 for i=1,…,m},\displaystyle\qquad\quad\in\hat{F}^{*}T\mathbb{R}^{2d}\oplus\pi_{M}^{*}TM\ |\ c_{i}^{3}=0\textrm{ for }i=1,\ldots,m\Big\},
F˘∗Gℝ2​d⊥={\displaystyle\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}=\{ (p,v)∈F˘∗​T​(ℝ2​d×M)\displaystyle(p,v)\in\breve{F}^{*}T(\mathbb{R}^{2d}\times M)
:F˘∗gp(v,w)=0,∀w∈F˘∗Gℝ2​d},\displaystyle:\breve{F}^{*}g_{p}(v,w)=0,\forall w\in\breve{F}^{*}G_{\mathbb{R}^{2d}}\},
PF˘∗​Gℝ2​d⊥T:T​M\displaystyle P_{\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}}^{T}:TM ⟶F˘∗​T​(ℝ2​d×M)⊗F˘∗​T∗​(ℝ2​d×M)\displaystyle\longrightarrow\breve{F}^{*}T(\mathbb{R}^{2d}\times M)\otimes\breve{F}^{*}T^{*}(\mathbb{R}^{2d}\times M)
(p,v)\displaystyle(p,v) ↦F˘∗​Ei∗​(vp)​(PF˘∗​Gℝ2​d⊥​(F˘∗​Ej​(vp)))\displaystyle\mapsto\breve{F}^{*}E^{*}_{i}(v_{p})\left(P_{\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}}(\breve{F}^{*}E_{j}(v_{p}))\right)
F˘∗​Ei​(vp)⊗F˘∗​Ej∗​(p),\displaystyle\hskip 16.0pt\breve{F}^{*}E_{i}(v_{p})\otimes\breve{F}^{*}E^{*}_{j}(p),
F˘∗​∇˘​PF˘∗​Gℝ2​d⊥T​(⋅,⋅):T​M→F˘∗​T​(ℝ2​d×M)⊗F˘∗​T∗​(ℝ2​d×M)\breve{F}^{*}\breve{\nabla}P_{\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}}^{T}(\cdot,\cdot):TM\rightarrow\breve{F}^{*}T(\mathbb{R}^{2d}\times M)\otimes\breve{F}^{*}T^{*}(\mathbb{R}^{2d}\times M)
F˘∗​∇˘​PF˘∗​Gℝ2​d⊥T​(⋅,V):T​M→F˘∗​T​(ℝ2​d×M).\breve{F}^{*}\breve{\nabla}P_{\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}}^{T}(\cdot,V):TM\rightarrow\breve{F}^{*}T(\mathbb{R}^{2d}\times M).
F˘∗​∇˘​PF˘∗​Gℝ2​d⊥T\displaystyle\breve{F}^{*}\breve{\nabla}P_{\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}}^{T} (⋅,V)=(𝟙{h>2​d,k≤2​d}​(h,k)\displaystyle(\cdot,V)=(\mathds{1}_{\{h>2d,k\leq 2d\}}(h,k)
−𝟙{h≤2​d,k>2​d}​(h,k))​F˘∗​Γ˘j​hk​Xj​Vh​𝔫k,\displaystyle-\mathds{1}_{\{h\leq 2d,k>2d\}}(h,k))\breve{F}^{*}\breve{\Gamma}^{k}_{jh}X^{j}V^{h}\mathfrak{n}_{k},
F˘∗​∇˘XF˘∗​Gℝ2​dV≜F˘∗∇˘XV+F˘∗∇PF˘∗​Gℝ2​d⊥T(⋅,V),\overset{\breve{F}^{*}G_{\mathbb{R}^{2d}}}{\breve{F}^{*}\breve{\nabla}_{X}}V\triangleq\breve{F}^{*}\breve{\nabla}_{X}V+\breve{F}^{*}\nabla P_{\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}}^{T}(\cdot,V),
F˘∗​Γ˘j​hkF˘∗​Gℝ2​d=(1+\displaystyle\overset{\breve{F}^{*}G_{\mathbb{R}^{2d}}}{\breve{F}^{*}\breve{\Gamma}^{k}_{jh}}=(1+ 𝟙{h>2​d,k≤2​d}​(h,k)\displaystyle\mathds{1}_{\{h>2d,k\leq 2d\}}(h,k) (50)
−𝟙{h≤2​d,k>2​d}​(h,k))​F˘∗​Γ˘j​hk.\displaystyle-\mathds{1}_{\{h\leq 2d,k>2d\}}(h,k))\breve{F}^{*}\breve{\Gamma}^{k}_{jh}.
F˘∗​∇˘\displaystyle\breve{F}^{*}\breve{\nabla} PF˘∗​Gℝ2​d⊥T​(⋅,V)​(F˘​(p,v))\displaystyle P_{\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\bot}}^{T}(\cdot,V)(\breve{F}(p,v))
=−∑k=2​d+12​d+m(∑j=12​m∑h=12​dF˘∗Γ˘j​hkXjVh)𝔫k∈F˘∗Gℝ2​d⊤.\displaystyle=-\sum^{2d+m}_{k=2d+1}\bigg(\sum^{2m}_{j=1}\sum^{2d}_{h=1}\breve{F}^{*}\breve{\Gamma}^{k}_{jh}X^{j}V^{h}\bigg)\mathfrak{n}_{k}\in\breve{F}^{*}G_{\mathbb{R}^{2d}}^{\top}.

II Multi-Task Acceleration Policy Optimization

The modified local dynamics and projected task accelerations form another weighted least-squares acceleration policy. The detailed Jacobians, curvature terms, and final expression are retained. The appendix states that analogous well-posedness and stability results could be developed but leaves their proofs outside its scope; these are not an additional completed theorem for every velocity-dependent task.

Equations 51–53
PBDSα(p,v)​{PF˘∗​G​(F˘∗​Dα(p,v)F˘∗​Gℝ2​d​(PF˘∗​G​(γα(p,v)​(s)))=PF˘∗​G​(F˘∗​ℱ˘​(γα(p,v)​(s))♯)γα(p,v)​(0)=F˘∗​d​F˘α(p,v)​(0)​(α(p,v)′​(0)),α(p,v)​(0)=(p,v),\textrm{PBDS}_{\alpha_{(p,v)}}\hskip-5.0pt\ \begin{cases}P_{\breve{F}^{*}G}\left(\overset{\breve{F}^{*}G_{\mathbb{R}^{2d}}}{\breve{F}^{*}D_{\alpha_{(p,v)}}}\left(P_{\breve{F}^{*}G}(\gamma_{\alpha_{(p,v)}}(s)\right)\right)\\ \quad=P_{\breve{F}^{*}G}\left(\breve{F}^{*}\breve{\mathcal{F}}(\gamma_{\alpha_{(p,v)}}(s))^{\sharp}\right)\\ \gamma_{\alpha_{(p,v)}}(0)=\breve{F}^{*}d\breve{F}_{\alpha_{(p,v)}(0)}(\alpha_{(p,v)}^{\prime}(0)),\\ \alpha_{(p,v)}(0)=(p,v),\end{cases}
𝒟VB,i={(((p,v),a),cj1∂xMj+cj2∂vM,1j+cj3∂aℝ2​dij\displaystyle\mathcal{D}_{\textrm{VB},i}=\Big\{\big(((p,v),a),c^{1}_{j}\partial x^{j}_{M}+c^{2}_{j}\partial v^{j}_{M,1}+c^{3}_{j}\partial a^{j}_{\mathbb{R}^{2d_{i}}}
+cj4∂vℝ2​dij+cj5∂vM,2j)∈T(F^i∗Tℝ2​di)⊕T(πM∗TM)\displaystyle\quad+c^{4}_{j}\partial v^{j}_{\mathbb{R}^{2d_{i}}}+c_{j}^{5}\partial v^{j}_{M,2}\big)\in T(\hat{F}_{i}^{*}T\mathbb{R}^{2d_{i}})\oplus T(\pi_{M}^{*}TM)
|cj1=cj2=0 for j=1,…,m,cj3=0 for j=1,…,d},\displaystyle\quad|\ c_{j}^{1}=c_{j}^{2}=0\textrm{ for }j=1,\ldots,m,c_{j}^{3}=0\textrm{ for }j=1,\ldots,d\Big\},
𝒟ℝ2​di={(((p,v),a),cj1∂xMj+cj2∂vM,1j+cj3∂aℝ2​dij\displaystyle\mathcal{D}_{\mathbb{R}^{2d_{i}}}=\Big\{\big(((p,v),a),c^{1}_{j}\partial x^{j}_{M}+c^{2}_{j}\partial v^{j}_{M,1}+c^{3}_{j}\partial a^{j}_{\mathbb{R}^{2d_{i}}}
+cj4∂vℝ2​dij+cj5∂vM,2j)∈T(F^i∗Tℝ2​di)⊕T(πM∗TM)\displaystyle\quad+c^{4}_{j}\partial v^{j}_{\mathbb{R}^{2d_{i}}}+c_{j}^{5}\partial v^{j}_{M,2}\big)\in T(\hat{F}_{i}^{*}T\mathbb{R}^{2d_{i}})\oplus T(\pi_{M}^{*}TM)
|cj5=0 for j=1,…,m}\displaystyle\quad|\ c_{j}^{5}=0\textrm{ for }j=1,\ldots,m\Big\}
Si:T​M\displaystyle S_{i}:TM ⟶T⁡(F˘i∗​T​(ℝ2​di×M))\displaystyle\longrightarrow T(\breve{F}_{i}^{*}T(\mathbb{R}^{2d_{i}}\times M))
(p,v)\displaystyle(p,v) ↦P𝒟ℝ2​di∘P𝒟VB,i​(γ(p,v)′​(0))\displaystyle\mapsto P_{\mathcal{D}_{\mathbb{R}^{2d_{i}}}}\circ P_{\mathcal{D}_{\textrm{VB},i}}(\gamma^{\prime}_{(p,v)}(0))
=(((p,v),a),(0,(0,γ(p,v)′a​(0),0))).\displaystyle=(((p,v),a),(0,(0,\gamma^{\prime a}_{(p,v)}(0),0))).
Zi:T​T​T​M\displaystyle Z_{i}:TTTM ⟶T⁡(F˘i∗​T​(ℝ2​di×M))\displaystyle\longrightarrow T(\breve{F}_{i}^{*}T(\mathbb{R}^{2d_{i}}\times M))
(((p,v),a),κ)\displaystyle(((p,v),a),\kappa) ↦P𝒟ℝ2​di∘P𝒟VB,i​(d​(F˘i∗​d​F˘i)((p,v),a)​(κ)).\displaystyle\mapsto P_{\mathcal{D}_{\mathbb{R}^{2d_{i}}}}\circ P_{\mathcal{D}_{\textrm{VB},i}}(d(\breve{F}_{i}^{*}d\breve{F}_{i})_{((p,v),a)}(\kappa)).
ζ:T​M\displaystyle\zeta:TM ⟶𝒟⊂T​T​T​M\displaystyle\longrightarrow\mathcal{D}\subset TTTM (51)
(p,v)\displaystyle(p,v) ↦arg⁡minκ∈𝒟(p,v)​∑i=1K12​∥Zi​(κ)−Si​(p,v)∥𝔉i∗​w˘i2,\displaystyle\mapsto\underset{\kappa\in\mathcal{D}_{(p,v)}}{\arg\min}\ \sum^{K}_{i=1}\frac{1}{2}\lVert Z_{i}(\kappa)-S_{i}(p,v)\rVert^{2}_{\mathfrak{F}_{i}^{*}\breve{w}_{i}},
𝑱⁡(𝑭˘i∗​𝒅​𝑭˘i)=[𝑰m0000𝑰m00𝑱​𝒇¯¨i𝑱​𝒇¯˙i𝑱​𝒇¯˙i𝑱​𝒇¯i𝑱​𝒇¯˙i0𝑱​𝒇¯i0000𝑰m],\bm{J}(\breve{\bm{F}}_{i}^{*}\bm{d}\breve{\bm{F}}_{i})=\begin{bmatrix}\bm{I}_{m}&0&0&0\\ 0&\bm{I}_{m}&0&0\\ \ddot{\bm{J}\bar{\bm{f}}}_{i}&\dot{\bm{J}\bar{\bm{f}}}_{i}&\dot{\bm{J}\bar{\bm{f}}}_{i}&\bm{J}\bar{\bm{f}}_{i}\\ \dot{\bm{J}\bar{\bm{f}}}_{i}&0&\bm{J}\bar{\bm{f}}_{i}&0\\ 0&0&0&\bm{I}_{m}\end{bmatrix},
∥Zi​(q)−Si​(p,v)∥𝔉i∗​w˘i2\displaystyle\lVert Z_{i}(q)-S_{i}(p,v)\rVert^{2}_{\mathfrak{F}_{i}^{*}\breve{w}_{i}} (52)
=∥𝑱​𝒇¯i​(p)​𝒂+𝑱​𝒇¯˙i​(p,v)​𝒗−𝜸˙vp,iκ​(0)∥𝒘˘ia​(𝔉i​(vp,0)))2,\displaystyle=\lVert\bm{J}\bar{\bm{f}}_{i}(p)\bm{a}+\dot{\bm{J}\bar{\bm{f}}}_{i}(p,v)\bm{v}-\dot{\bm{\gamma}}^{\kappa}_{v_{p},i}(0)\rVert^{2}_{\breve{\bm{w}}_{i}^{a}(\mathfrak{F}_{i}(v_{p},0)))},
(PF˘i∗​G\displaystyle\Big(P_{\breve{F}_{i}^{*}G} (F˘i∗​ℱ˘i​(γvp,i​(0))♯))j\displaystyle\Big(\breve{F}_{i}^{*}\breve{\mathcal{F}}_{i}(\gamma_{v_{p},i}(0))^{\sharp}\Big)\Big)^{j}
=gℝ2​dij​h​(F^i​(vp))​(ℱ˘D,ih​(γvp,i​(0))−∂Φ˘i∂xh​(F˘i​(vp))).\displaystyle=g^{jh}_{\mathbb{R}^{2d_{i}}}(\hat{F}_{i}(v_{p}))\bigg(\breve{\mathcal{F}}_{D,i}^{h}(\gamma_{v_{p},i}(0))-\frac{\partial\breve{\Phi}_{i}}{\partial x^{h}}(\breve{F}_{i}(v_{p}))\bigg).
PF˘i∗​G​(α˙vp,iℓ​(0)​PF˘i∗​G​(γvp,i​(0))h​F˘i∗​Γ˘ℓ​hj−2​m​(vp))\displaystyle P_{\breve{F}_{i}^{*}G}\Big(\dot{\alpha}_{v_{p},i}^{\ell}(0)P_{\breve{F}_{i}^{*}G}\Big(\gamma_{v_{p},i}(0)\Big)^{h}\breve{F}_{i}^{*}\breve{\Gamma}^{j-2m}_{\ell h}(v_{p})\Big)
=PF˘i∗​G​(α˙vp,iℓ​(0)​PF˘i∗​G​((F˘i∗​J​F˘i)h​r​α˙vp,ir​(0))\displaystyle=P_{\breve{F}_{i}^{*}G}\Big(\dot{\alpha}_{v_{p},i}^{\ell}(0)P_{\breve{F}_{i}^{*}G}\Big((\breve{F}_{i}^{*}J\breve{F}_{i})_{hr}\dot{\alpha}_{v_{p},i}^{r}(0)\Big)
(J​F˘i)s​ℓ​(vp)​Γ˘s​hj−2​m​(F˘i​(vp)))\displaystyle\quad(J\breve{F}_{i})_{s\ell}(v_{p})\breve{\Gamma}^{j-2m}_{sh}(\breve{F}_{i}(v_{p}))\Big)
=(α˙vp,iv)ℓ​(0)​(J​f¯i)h​r​(p)​(α˙vp,iv)r​(0)\displaystyle=(\dot{\alpha}_{v_{p},i}^{v})^{\ell}(0)(J\bar{f}_{i})_{hr}(p)(\dot{\alpha}_{v_{p},i}^{v})^{r}(0)
(J​F^i)s​ℓ​(vp)​Γ˘s⁡(h+d)j−2​m​(F^i​(vp))\displaystyle\hskip 30.0pt(J\hat{F}_{i})_{s\ell}(v_{p})\breve{\Gamma}^{j-2m}_{s(h+d)}(\hat{F}_{i}(v_{p}))
+(α˙vp,ia)ℓ​(0)​(J​f¯i)h​r​(p)​(α˙vp,iv)r​(0)\displaystyle+(\dot{\alpha}_{v_{p},i}^{a})^{\ell}(0)(J\bar{f}_{i})_{hr}(p)(\dot{\alpha}_{v_{p},i}^{v})^{r}(0)
(J​F^i)s⁡(ℓ+d)​(vp)​Γ˘s⁡(h+d)j−2​m​(F^i​(vp))\displaystyle\hskip 30.0pt(J\hat{F}_{i})_{s(\ell+d)}(v_{p})\breve{\Gamma}^{j-2m}_{s(h+d)}(\hat{F}_{i}(v_{p}))
=(Ξiv)(j−2​m−di)​ℓ​(vp)​(α˙vp,iv)ℓ​(0)\displaystyle=(\Xi^{v}_{i})_{(j-2m-d_{i})\ell}(v_{p})(\dot{\alpha}_{v_{p},i}^{v})^{\ell}(0)
+(Ξia)(j−2​m−di)​ℓ​(vp)​(α˙vp,ia)ℓ​(0),\displaystyle\qquad+(\Xi^{a}_{i})_{(j-2m-d_{i})\ell}(v_{p})(\dot{\alpha}_{v_{p},i}^{a})^{\ell}(0),
𝑭˘i∗​𝑱​𝑭˘i=[𝑱​𝒇¯˙i𝑱​𝒇¯i0𝑱​𝒇¯i0000𝑰m].\breve{\bm{F}}_{i}^{*}\bm{J}\breve{\bm{F}}_{i}=\begin{bmatrix}\dot{\bm{J}\bar{\bm{f}}}_{i}&\bm{J}\bar{\bm{f}}_{i}&0\\ \bm{J}\bar{\bm{f}}_{i}&0&0\\ 0&0&\bm{I}_{m}\end{bmatrix}.
𝜻(p,v)=(∑i=1K\displaystyle\bm{\zeta}(p,v)=\Bigg(\sum_{i=1}^{K} 𝓒i⊤​(p,v)​𝒘˘ia​(𝔉i​(vp,0))​𝓒i​(p,v))†\displaystyle\bm{\mathcal{C}}_{i}^{\top}(p,v)\breve{\bm{w}}_{i}^{a}(\mathfrak{F}_{i}(v_{p},0))\bm{\mathcal{C}}_{i}(p,v)\Bigg)^{\dagger}
(∑i=1K𝓒i⊤​(p,v)​𝒘˘ia​(𝔉i​(vp,0))​𝓐i​(p,v))\displaystyle\left(\sum_{i=1}^{K}\bm{\mathcal{C}}_{i}^{\top}(p,v)\breve{\bm{w}}_{i}^{a}(\mathfrak{F}_{i}(v_{p},0))\bm{\mathcal{A}}_{i}(p,v)\right)
𝓒i​(p,v)=𝑱​𝒇¯i​(p)+𝚵iv​(p,v)\quad\bm{\mathcal{C}}_{i}(p,v)=\bm{J}\bar{\bm{f}}_{i}(p)+\bm{\Xi}_{i}^{v}(p,v)\vskip-2.0pt
𝓐𝒊​(p,v)\displaystyle\bm{\mathcal{A}_{i}}(p,v) =(𝒈ℝ2​di−1)a​(F^i​(vp))​(𝓕D,ℝ2​di​((d​F^i)vp​(vp,0))\displaystyle=(\bm{g}_{\mathbb{R}^{2d_{i}}}^{-1})^{a}(\hat{F}_{i}(v_{p}))\Big(\bm{\mathcal{F}}_{D,\mathbb{R}^{2d_{i}}}((d\hat{F}_{i})_{v_{p}}(v_{p},0))
−∇𝚽ℝ2​di​(F^i​(vp)))−(𝑱​𝒇¯˙i​(p,v)−𝚵ia​(p,v))​𝒗,\displaystyle-\nabla\bm{\Phi}_{\mathbb{R}^{2d_{i}}}(\hat{F}_{i}(v_{p}))\Big)-(\dot{\bm{J}\bar{\bm{f}}}_{i}(p,v)-\bm{\Xi}_{i}^{a}(p,v))\bm{v},
(Ξiv)j​ℓ​(p,v)=(J​F^i)s​ℓ​(vp)​(Γℝ2​di)s⁡(h+m)j+di​(F^i​(vp))​J​f¯h​r​(p)​vr(Ξia)j​ℓ​(p,v)=(J​F^i)s⁡(ℓ+di)​(vp)​(Γℝ2​di)s⁡(h+m)j+di​(F^i​(vp))​J​f¯h​r​(p)​vr.\begin{gathered}(\Xi^{v}_{i})_{j\ell}(p,v)=(J\hat{F}_{i})_{s\ell}(v_{p})(\Gamma_{\mathbb{R}^{2d_{i}}})^{j+d_{i}}_{s(h+m)}(\hat{F}_{i}(v_{p}))J\bar{f}_{hr}(p)v^{r}\\ (\Xi^{a}_{i})_{j\ell}(p,v)=(J\hat{F}_{i})_{s(\ell+d_{i})}(v_{p})(\Gamma_{\mathbb{R}^{2d_{i}}})^{j+d_{i}}_{s(h+m)}(\hat{F}_{i}(v_{p}))J\bar{f}_{hr}(p)v^{r}.\end{gathered}
{σ¨​(t)=ζ​(σ˙​(t))=arg⁡minκ∈𝒟σ˙​(t)​∑i=1K12​∥Zi​(κ)−Si​(σ˙​(t))∥𝔉i∗​w˘i2σ˙​(0)=(p0,v0).\begin{cases}\ddot{\sigma}(t)=\zeta(\dot{\sigma}(t))=\\ \qquad\underset{\kappa\in\mathcal{D}_{\dot{\sigma}(t)}}{\arg\min}\ \sum^{K}_{i=1}\frac{1}{2}\lVert Z_{i}(\kappa)-S_{i}(\dot{\sigma}(t))\rVert^{2}_{\mathfrak{F}_{i}^{*}\breve{w}_{i}}\\ \dot{\sigma}(0)=(p_{0},v_{0}).\end{cases}\hskip-10.0pt (53)
Concept demonstration

Open image ↗