betaFeedback

WorksComputer Science

Geometry-Aware Humanoid Retargeting via Manifold ODEs

Abstract

In this paper, we focus on interaction-preserving motion retargeting, which transfers human demonstrations to humanoids while maintaining spatial relationships with objects and terrain. Previous work, such as OmniRetarget, addresses this task through temporally coupled constrained optimization, requiring repeated sequential quadratic programming (SQP) iterations at each frame. We study this problem from an ordinary differential equation (ODE) perspective by modeling the continuous-time limit of its framewise Karush-Kuhn-Tucker (KKT) conditions as an ODE on the robot configuration manifold. This viewpoint offers two benefits: (i) it connects OmniRetarget's SQP iterations to a backward discretization of the ODE, enabling us to establish their local convergence under suitable regularity assumptions; and (ii) it leads to TaSP-KKT (tangent-space parametrization for KKT flow), a new retargeting method based on a forward discretization of the KKT flow. The proposed method achieves the same first-order convergence to the ODE solution as the corresponding backward scheme while requiring only one weighted tangent-cone projection and a closed-form manifold update per step, thereby avoiding repeated inner SQP iterations. Numerical experiments show that TaSP-KKT achieves retargeting quality comparable to, though slightly lower than, that of OmniRetarget while running approximately twice as fast.

Introduction

Whole-body tasks such as carrying objects and stepping onto a platform require humanoid robots to coordinate their limbs while maintaining contact with objects and the environment [1, 2]. Human demonstrations provide examples of this coordination and can serve as references for learning robot behaviors. However, differences in body proportions, joint structure, and motion limits prevent human motions from being transferred directly to a robot. Generating usable robot motion references therefore requires adapting both the demonstrated movements and their interactions with the scene to the robot’s embodiment. Motion retargeting aims to enable a target character or robot to reproduce a source motion despite differences in body proportions or skeletal structure while preserving the motion’s essential characteristics.

Early approaches to motion retargeting include the space-time optimization of Gleicher [3] and the inverse-rate formulation of Choi and Ko [4]. More recently, General Motion Retargeting (GMR) [5] combines local scaling with differential inverse kinematics to generate humanoid motion references.

Interaction-preserving motion retargeting, which maintains spatial relationships and contacts during motion transfer, has also received much attention. For example, Contact-Aware Retargeting [6] combines a geometry-conditioned recurrent network with encoder-space optimization to reduce interpenetration and preserve self-contacts and ground contacts. M2M^2RET [7] uses a learned model to predict residual motion corrections that account for skeletal relationships and body geometry, aiming to reduce interpenetration and missing contacts when transferring motion between characters with different skeletal structures.

Among the most recent work is OmniRetarget [8], which uses an interaction mesh to represent spatial relationships among the body, objects, and terrain. At each frame, it minimizes Laplacian mesh deformation subject to kinematic constraints using a sequential quadratic programming (SQP) solver. The resulting references reduced penetration and foot skating and supported object carrying and terrain traversal on a Unitree G1 humanoid.

In this paper, we study humanoid motion retargeting and OmniRetarget from the perspective of ordinary differential equations (ODEs). Our motivation is twofold. First, OmniRetarget solves a sequence of temporally coupled constrained optimization problems, but the original work does not establish convergence guarantees for its SQP iterations. The longstanding connections between optimization and ODEs provide tools for analyzing convergence [9, 10, 11]. These connections motivate us to establish theoretical guarantees for OmniRetarget through a dynamical formulation.

Second, an ODE formulation can guide the development of new algorithms. In deep learning, Haber et al. [12] use a continuous-time interpretation of residual networks to develop multiscale training methods, which attains higher classification accuracy and lower training cost. In diffusion generative modeling, Lu et al. [13] exploit the structure of diffusion ODEs to derive DPM-Solver, which use with high-order ODE schemes to accelerate sampling. These examples motivate using the structure of retargeting dynamics to design more efficient numerical methods.

ODE-based methods have also been explored in humanoid planning and retargeting. Klein et al. [14] generate robot arm motions as piecewise geodesics governed by second-order ODEs under an inertia-based Riemannian metric. Their subsequent work [15] modifies this metric using barrier functions to generate collision-avoiding trajectories. Li et al. [16] use a learned conditional ODE flow to adapt motions across characters with the same skeletal topology but different body proportions. MoReFlow [17] formulates human-to-robot motion mapping as an ODE flow in latent space and uses flow matching to train a neural network that parameterizes its velocity field.

In this paper, we model the evolution of the optimal robot configurations generated by OmniRetarget as an ODE on M=S3×Rd\mathcal{M} = \mathbb{S}^{3} \times\mathbb{R}^{d}, where S3\mathbb{S}^{3} represents the floating-base orientation using a unit quaternion, and Rd\mathbb{R}^{d} contains the base translation and joint coordinates. We then use this ODE formulation to analyze the convergence of OmniRetarget’s SQP solver and derive a computationally efficient retargeting algorithm. Our contributions are summarized as follows:

  1. We develop a manifold ODE formulation of OmniRetarget. In particular, we characterize the evolution of OmniRetarget’s framewise KKT point as a projected flow on the robot configuration manifold M=S3×Rd\mathcal{M} = \mathbb{S}^{3} \times\mathbb{R}^{d}. The formulation provides a common basis for analyzing the existing solver and designing new retargeting algorithms.

  1. We connect OmniRetarget’s SQP iterations to a backward discretization of the KKT flow. Through this connection, we interpret the SQP iterations as Newton steps for the backward discretization and establish local convergence of the SQP solver.

  1. We propose TaSP-KKT, a retargeting algorithm based on a forward discretization of the KKT flow. It achieves the same convergence order as the SQP-based backward scheme. Each step requires only one weighted tangent-cone projection and a closed-form manifold update, avoiding repeated inner SQP iterations and thus reducing the computational cost. Preliminary numerical experiments show that TaSP-KKT achieves retargeting quality comparable to that of OmniRetarget while running approximately twice as fast.

Preliminaries

Ordinary Differential Equations

An ODE describes the evolution of a state y(τ)∈Rny(\tau) \in\mathbb{R}^{n} over continuous time τ\tau. At each time τ\tau, the vector field F(τ,y(τ))F(\tau, y(\tau)) assigns a rate of change to the current state. Together with an initial state y0y_{0}, the ODE defines the initial-value problem

y˙(τ)=F(τ,y(τ)),y(0)=y0,(1)\dot{y}(\tau) = F(\tau, y(\tau)), \qquad y(0) = y_{0}, \tag*{(1)}

where y˙(τ)=dy(τ)/dτ\dot{y}(\tau) = \mathrm{d}y(\tau)/\mathrm{d}\tau. A solution of (1) on [0,T][0,T] is a trajectory y:[0,T]→Rny : [0,T] \to\mathbb{R}^{n} that starts at y0y_{0} and whose derivative agrees with the vector field at every time.

a) Discrete approximation: To approximate the continuous-time trajectory numerically, we discretize time and advance the state in finite steps. For a step size η>0\eta> 0 and sampling times τk=kη\tau_{k} = k\eta, a one-step method generates approximations yk≈y(τk)y_{k} \approx y(\tau_{k}) according to

yk+1=Ψη(τk,yk),y0=y(0),y_{k+1} = \Psi_{\eta}(\tau_{k}, y_{k}), \qquad y_{0} = y(0),

where Ψη\Psi_{\eta} denotes the numerical update map. For example, the forward and backward Euler methods are given by

Forward Euler:yk+1=yk+ηF(τk,yk),Backward Euler:yk+1=yk+ηF(τk+1,yk+1).(2)\begin{aligned} \text{Forward Euler:}\quad& y_{k+1} = y_{k} + \eta F(\tau_{k}, y_{k}),\\ \text{Backward Euler:}\quad& y_{k+1} = y_{k} + \eta F(\tau_{k+1}, y_{k+1}). \tag*{(2)} \end{aligned}

Forward Euler evaluates the vector field at the known current state and thus provides an explicit update. Backward Euler evaluates the vector field at the unknown next state and therefore requires solving an implicit equation, which is generally nonlinear [18].

b) Convergence order: Convergence order quantifies how rapidly the numerical error decreases as the step size η\eta tends to zero over a fixed time interval [0,T][0,T]. Let nmax⁡=⌊T/η⌋n_{\max} = \lfloor T/\eta\rfloor. A method has global convergence order p≥1p \ge1 if, for all sufficiently small η\eta,

max⁡0≤k≤nmax⁡∥yk−y(τk)∥2≤CTηp,\max_{0 \le k \le n_{\max}} \left\lVert y_{k} - y(\tau_{k}) \right\rVert_{2} \le C_{T}\eta^{p},

where the constant CTC_{T} may depend on TT but is independent of η\eta. Under standard smoothness assumptions, both Euler methods have first-order convergence [18].

ODEs on Robot Configuration Space

We represent the configuration of a floating-base robot as

q=(r,x)∈M:=S3×Rd,S3:={r∈R4:∥r∥2=1}.(3)q = (r,x) \in\mathcal{M} := \mathbb{S}^{3} \times\mathbb{R}^{d}, \qquad \mathbb{S}^{3} := \{r \in\mathbb{R}^{4} : \lVert r \rVert_{2} = 1\}. \tag*{(3)}

The unit quaternion r=(r0,r1,r2,r3)Tr = (r_{0}, r_{1}, r_{2}, r_{3})^{\mathsf{T}} represents the base orientation, while xx contains the base translation and joint coordinates. Quaternion signs are chosen consistently across frames. The tangent space TqMT_{q}\mathcal{M} consists of the velocities of smooth curves on M\mathcal{M} passing through qq. Differentiating the constraint rTr=1r^{\mathsf{T}}r = 1 shows that a quaternion velocity ξr\xi_{r} must satisfy rTξr=0r^{\mathsf{T}}\xi_{r} = 0. Thus, a tangent quaternion vector is perpendicular to rr, and the tangent space of M\mathcal{M} at qq is given by

TqM=TrS3×Rd,TrS3={ξr∈R4:rTξr=0}.(4)T_{q}\mathcal{M} = T_{r}\mathbb{S}^{3} \times\mathbb{R}^{d}, \qquad T_{r}\mathbb{S}^{3} = \{\xi_{r} \in\mathbb{R}^{4} : r^{\mathsf{T}}\xi_{r} = 0\}. \tag*{(4)}

An ODE on the configuration manifold must generate tangent velocities [19 Ch. IV]:

q˙(τ)=V(τ,q(τ)),V(τ,q)∈TqM.(5)\dot{q}(\tau) = V(\tau, q(\tau)), \qquad V(\tau,q) \in T_{q}\mathcal{M}. \tag*{(5)}

Equivalently,

r˙=Vr(τ,r,x),x˙=Vx(τ,r,x),rTVr=0.(6)\dot{r} = V_{r}(\tau,r,x), \qquad\dot{x} = V_{x}(\tau,r,x), \qquad r^{\mathsf{T}}V_{r} = 0. \tag*{(6)}

Consider two unit quaternions r,r′∈S3r,r' \in\mathbb{S}^{3} in the same hemisphere. Let v∈TrS3v \in T_{r}\mathbb{S}^{3} denote the orthogonal projection of their displacement onto the tangent space at rr:

v:=(I−rrT)(r′−r)=(I−rrT)r′.(7)v := (I - rr^{\mathsf{T}})(r' - r) = (I - rr^{\mathsf{T}})r'. \tag*{(7)}

The point r′r' can therefore be decomposed as r′=αr+vr' = \alpha r + v, where α∈R\alpha\in\mathbb{R}. Since rTv=0r^{\mathsf{T}}v = 0 and r,r′r,r' have unit norm,

1=∥r′∥22=α2+∥v∥22.(8)1 = \lVert r' \rVert_{2}^{2} = \alpha^{2} + \lVert v \rVert_{2}^{2}. \tag*{(8)}

Selecting the positive root gives α=1−∥v∥22\alpha= \sqrt{1-\lVert v\rVert_{2}^{2}}. Thus, for any v∈TrS3v \in T_{r}\mathbb{S}^{3} with ∥v∥2<1\lVert v\rVert_{2} < 1, the tangent projection uniquely determines r′r' within the hemisphere through the tangent space parametrization (TaSP) map [20]:

r′=TaSP⁡r(v):=1−∥v∥22 r+v.r' = \operatorname{TaSP}_{r}(v) := \sqrt{1-\lVert v\rVert_{2}^{2}}\,r + v.

Combining this parametrization with forward Euler yields the TaSP–Euler method for (6). Starting from q0=q(0)q_0=q(0), we set vk=ηVr(τk,qk)v_k=\eta V_r(\tau_k,q_k) and compute qk+1=(rk+1,xk+1)q_{k+1}=(r_{k+1},x_{k+1}) using

rk+1=TaSP⁡rk(ηVr(τk,qk)),xk+1=xk+ηVx(τk,qk).(9)r_{k+1}=\operatorname{TaSP}_{r_k}\left(\eta V_r(\tau_k,q_k)\right), \qquad x_{k+1}=x_k+\eta V_x(\tau_k,q_k). \tag*{(9)}

The update preserves qk+1∈Mq_{k+1}\in\mathcal{M} when η∥Vr(τk,qk)∥2<1\eta\lVert V_r(\tau_k,q_k)\rVert_2<1. The forward TaSP–Euler method has first-order convergence [18], i.e.,

max⁡0≤k≤⌊T/η⌋∥qk−q(τk)∥2≤CTη.(10)\max_{0\leq k\leq\lfloor T/\eta\rfloor}\lVert q_k-q(\tau_k)\rVert_2\leq C_T\eta. \tag*{(10)}

A gradient flow for a smooth function f:M→Rf:\mathcal{M}\to\mathbb{R} evolves along the negative Riemannian gradient of ff:

q˙(τ)=−Dqf(q(τ)),q(0)=q0,(11)\dot{q}(\tau)=-\mathcal{D}_qf(q(\tau)),\qquad q(0)=q_0, \tag*{(11)}

where Dqf\mathcal{D}_qf denotes the Riemannian gradient, which is obtained by projecting the Euclidean gradient onto TqMT_q\mathcal{M}. For q=(r,x)q=(r,x), the orthogonal projector is

Πq=diag⁡(I4−rrT,Id),\Pi_q=\operatorname{diag}(I_4-rr^{\mathsf{T}},I_d),

which gives

Dqf=Πq∇f=[(I4−rrT)∇rf∇xf]=[grad⁡rf∇xf].(12)\mathcal{D}_qf=\Pi_q\nabla f =\begin{bmatrix}(I_4-rr^{\mathsf{T}})\nabla_rf\\ \nabla_xf\end{bmatrix} =\begin{bmatrix}\operatorname{grad}_rf\\ \nabla_xf\end{bmatrix}. \tag*{(12)}

Here, ∇f\nabla f denotes the Euclidean gradient, and grad⁡rf:=(I4−rrT)∇rf\operatorname{grad}_rf:=(I_4-rr^{\mathsf{T}})\nabla_rf is its Riemannian gradient with respect to the quaternion component.

SQP-based Motion Retargeting

OmniRetarget [8] transfers human motion to a robot by preserving the spatial relationships among the body, manipulated objects, and terrain. Consider a human demonstration over the time interval [0,T][0,T], uniformly sampled with a sampling period δ>0\delta>0. The sampled demonstration is represented by a sequence {Pt}t=0M\{P_t\}_{t=0}^{M}, where Pt=[pt,1,…,pt,N]T∈RN×3P_t=[p_{t,1},\ldots,p_{t,N}]^{\mathsf{T}}\in\mathbb{R}^{N\times3} contains the 3D positions of human body keypoints and points sampled from object and terrain surfaces. At each sample tt, OmniRetarget constructs an interaction mesh by applying Delaunay tetrahedralization [21] to PtP_t. Let Lt∈RN×NL_t\in\mathbb{R}^{N\times N} denote the associated graph Laplacian [22], defined such that the iith row of LtPtL_tP_t represents the displacement of vertex ii from the mean position of its neighbors [8].

Let qt∈Mq_t\in\mathcal{M} denote the retargeted robot configuration at frame tt. The map Φ:M→RN×3\Phi:\mathcal{M}\to\mathbb{R}^{N\times3} returns the corresponding target mesh vertices by combining robot keypoints computed through forward kinematics with prescribed object and terrain points. At each frame, OmniRetarget obtains the desired robot configuration by solving the following constrained optimization problem

qt⋆∈arg⁡min⁡q∈Mft(q)+12δ∥q−qt−1∥Q2q_t^\star\in\underset{q\in\mathcal{M}}{\arg\min}\quad f_t(q)+\frac{1}{2\delta}\lVert q-q_{t-1}\rVert_Q^2
s.t.gt,i(q)≤0,i=1,…,m,\text{s.t.}\quad g_{t,i}(q)\leq0,\quad i=1,\ldots,m,
ht,j(q)=0,j=1,…,ℓ.(13)h_{t,j}(q)=0,\quad j=1,\ldots,\ell. \tag*{(13)}

The objective and constraints have the following roles:

  • Interaction preservation. The loss is defined as

ft(q):=∥et(q)∥22,et(q):=vec⁡(Lt[Φ(q)−Pt]).(14)f_t(q):=\lVert e_t(q)\rVert_2^2, \qquad e_t(q):=\operatorname{vec}\left(L_t[\Phi(q)-P_t]\right). \tag*{(14)}

where vec⁡\operatorname{vec} stacks the columns of a matrix. The residual compares corresponding vertices’ offsets from their neighbors’ average positions in the robot and human interaction meshes. Minimizing ftf_t therefore encourages preservation of local spatial relationships among the body, objects, and terrain.

  • Temporal regularization. With z:=q−qt−1z:=q-q_{t-1}, define ∥z∥Q2:=zTQz\lVert z\rVert_Q^2:=z^{\mathsf{T}}Qz, where Q∈R(d+4)×(d+4)Q\in\mathbb{R}^{(d+4)\times(d+4)} is symmetric positive semidefinite. This penalty discourages large changes between consecutive frames, encouraging temporal smoothness [8].

  • Kinematic feasibility. At time step tt, the vector gt(q)=[gt,1(q),…,gt,m(q)]T∈Rmg_t(q)=[g_{t,1}(q),\ldots,g_{t,m}(q)]^{\mathsf{T}}\in\mathbb{R}^m collects the mm inequality constraints enforcing non-penetration and joint position and velocity limits. Similarly, ht(q)=[ht,1(q),…,ht,ℓ(q)]T∈Rℓh_t(q)=[h_{t,1}(q),\ldots,h_{t,\ell}(q)]^{\mathsf{T}}\in\mathbb{R}^{\ell} collects the ℓ\ell equality constraints, including those specifying stance-foot positions. The subscript tt indicates the time dependence of these constraint functions.

Note that our temporal penalty includes an explicit factor 1/(2δ)1/(2\delta), which is absent from the original formulation [8]. The two expressions differ only in their normalization and are equivalent under an appropriate rescaling of the weight matrix QQ. OmniRetarget solves (13) using SQP. At frame tt, the iterations are initialized with the SQP solution from frame t−1t-1, i.e., qˉt,0=qt−1SQP\bar{q}_{t,0}=q^{\mathrm{SQP}}_{t-1}. At iteration i≥1i\geq1, the SQP process linearizes the interaction residual and constraint functions at qˉt,i−1\bar{q}_{t,i-1} to obtain the following trust-region subproblem for the increment Δqt,i∈Rd+4\Delta q_{t,i}\in\mathbb{R}^{d+4}

min⁡Δq∥et(qˉt,i−1)+Jt,i−1eΔq∥22+12δ∥qˉt,i−1+Δq−qt−1SQP∥Q2\underset{\Delta q}{\min}\quad \left\lVert e_t(\bar{q}_{t,i-1})+J^{e}_{t,i-1}\Delta q\right\rVert_2^2 +\frac{1}{2\delta}\left\lVert\bar{q}_{t,i-1}+\Delta q-q^{\mathrm{SQP}}_{t-1}\right\rVert_Q^2
s.t.gt,j(qˉt,i−1)+Jt,i−1gjΔq≤0,j=1,…,m,\text{s.t.}\quad g_{t,j}(\bar{q}_{t,i-1})+J^{g_j}_{t,i-1}\Delta q\leq0,\quad j=1,\ldots,m,
ht,j(qˉt,i−1)+Jt,i−1hjΔq=0,j=1,…,ℓ,h_{t,j}(\bar{q}_{t,i-1})+J^{h_j}_{t,i-1}\Delta q=0,\quad j=1,\ldots,\ell,
Δq∈Tqˉt,i−1M,∥Δq∥2≤ϵ.(15)\Delta q\in T_{\bar{q}_{t,i-1}}\mathcal{M},\qquad\lVert\Delta q\rVert_2\leq\epsilon. \tag*{(15)}

Here, Jt,i−1eJ^e_{t,i-1} denotes the Jacobian of the residual ete_t, while Jt,i−1gjJ^{g_j}_{t,i-1} and Jt,i−1hjJ^{h_j}_{t,i-1} denote the row Jacobians of the scalar constraints gt,jg_{t,j} and ht,jh_{t,j}, respectively. All Jacobians are evaluated with respect to qq at qˉt,i−1\bar{q}_{t,i-1}, with the frame data PtP_t held fixed.

After solving (15), we decompose the resulting increment into its quaternion and Euclidean components, Δqt,i=(Δrt,i,Δxt,i)\Delta q_{t,i}=(\Delta r_{t,i},\Delta x_{t,i}). The next iterate qˉt,i=(rˉt,i,xˉt,i)\bar{q}_{t,i}=(\bar{r}_{t,i},\bar{x}_{t,i}) is obtained by adding the increment to the current iterate and normalizing the quaternion component:

rˉt,i=rˉt,i−1+Δrt,i∥rˉt,i−1+Δrt,i∥2,xˉt,i=xˉt,i−1+Δxt,i.(16)\bar{r}_{t,i}=\frac{\bar{r}_{t,i-1}+\Delta r_{t,i}}{\lVert\bar{r}_{t,i-1}+\Delta r_{t,i}\rVert_2}, \qquad \bar{x}_{t,i}=\bar{x}_{t,i-1}+\Delta x_{t,i}. \tag*{(16)}

After NN SQP iterations, the output for frame tt is defined as qtSQP:=qˉt,Nq_t^{\mathrm{SQP}}:=\bar{q}_{t,N}.

KKT Flow of Humanoid Motion Retargeting

In this section, we model the evolution of the framewise KKT points of (13) as a constrained gradient ODE on M\mathcal{M} in the limit as the sampling period δ→0\delta\to0. As a result, starting from a KKT point at the initial frame, one can integrate the proposed dynamics to track the corresponding KKT trajectory, thereby avoiding the need to solve a separate optimization problem at each subsequent frame. Set τt=tδ\tau_t=t\delta and write ft=f(τt,⋅)f_t=f(\tau_t,\cdot), gt=g(τt,⋅)g_t=g(\tau_t,\cdot), and ht=h(τt,⋅)h_t=h(\tau_t,\cdot). Define the feasible set and the active inequality index set as

Ft:={q∈M:gt(qt)≤0, ht(qt)=0},\mathcal{F}_t:=\{q\in\mathcal{M}:g_t(q_t)\le0,\ h_t(q_t)=0\},
I(qt):={i∈{1,…,m}:gt,i(qt)=0}.\mathcal{I}(q_t):=\{i\in\{1,\ldots,m\}:g_{t,i}(q_t)=0\}.

At q∈Ftq\in\mathcal{F}_t, write I=I(qt)\mathcal{I}=\mathcal{I}(q_t) and define the active constraint Jacobian

A(qt):=[Dqgt,I(qt)Dqht(qt)]=[∇qgt,I(qt)∇qht(qt)]Πqt.(17)A(q_t):= \begin{bmatrix} \mathfrak{D}_qg_{t,\mathcal{I}}(q_t)\\ \mathfrak{D}_qh_t(q_t) \end{bmatrix} = \begin{bmatrix} \nabla_qg_{t,\mathcal{I}}(q_t)\\ \nabla_qh_t(q_t) \end{bmatrix} \Pi_{q_t}. \tag*{(17)}

Write NC(qt)\mathcal{N}\mathcal{C}(q_t) and TC(qt)\mathcal{T}\mathcal{C}(q_t) for the outward normal and tangent cones to Ft\mathcal{F}_t relative to M\mathcal{M}. At a feasible point qq, these cones admit the representations [23]

NC(qt)={A(qt)⊤[λμ]:λ∈R+∣I∣, μ∈Rℓ},(18)\mathcal{N}\mathcal{C}(q_t)=\left\{A(q_t)^{\top} \begin{bmatrix} \lambda\\ \mu \end{bmatrix}:\lambda\in\mathbb{R}_{+}^{|\mathcal{I}|},\ \mu\in\mathbb{R}^{\ell}\right\}, \tag*{(18)}
TC(qt)={w∈TqM:⟨n,w⟩≤0 for all n∈NC(qt)}.(19)\mathcal{T}\mathcal{C}(q_t)=\{w\in T_q\mathcal{M}:\langle n,w\rangle\le0\text{ for all }n\in\mathcal{N}\mathcal{C}(q_t)\}. \tag*{(19)}

Since TC(qt)\mathcal{T}\mathcal{C}(q_t) is a nonempty closed convex cone, any matrix Q=Q⊤≻0Q=Q^{\top}\succ0 defines a unique weighted projection

Pt,qQ(v):=arg⁡min⁡w∈TC(qt) 12∥w−v∥Q2.(20)\mathcal{P}_{t,q}^{Q}(v):=\underset{w\in\mathcal{T}\mathcal{C}(q_t)}{\arg\min}\ \frac{1}{2}\lVert w-v\rVert_{Q}^{2}. \tag*{(20)}

Given any q∈Ftq\in\mathcal{F}_t at which A(qt)A(q_t) has full row rank, any v∈Rd+4v\in\mathbb{R}^{d+4}, any w∈TC(qt)w\in\mathcal{T}\mathcal{C}(q_t), the projection is characterized by

Pt,qQ(v)=w⟺{ΠqQ(v−w)∈NC(qt),⟨Q(v−w),w⟩=0.(21)\mathcal{P}_{t,q}^{Q}(v)=w\Longleftrightarrow \begin{cases} \Pi_qQ(v-w)\in\mathcal{N}\mathcal{C}(q_t),\\ \langle Q(v-w),w\rangle=0. \end{cases} \tag*{(21)}

For analysis, we make the following regularity and the linear independence constraint qualification (LICQ) assumptions [18, 23]. In Assumptions 1–3, φt\varphi_t denotes any of the scalar functions ftf_t, gt,ig_{t,i} (i=1,…,mi=1,\ldots,m), or ht,jh_{t,j} (j=1,…,ℓj=1,\ldots,\ell).

Assumption 1 (Boundedness): There exists a constant B>0B>0, independent of tt, such that

∣φt(q)∣≤B,∀q∈Ft,(22)|\varphi_t(q)|\le B,\qquad\forall q\in\mathcal{F}_t, \tag*{(22)}

holds for every frame tt and every function φt\varphi_t.

Assumption 2 (Lipschitz continuity): There exists a constant K>0K>0, independent of tt, such that

∣φt(q)−φt(q′)∣≤K∥q−q′∥2,∀q,q′∈Ft,|\varphi_t(q)-\varphi_t(q')|\le K\lVert q-q'\rVert_2,\qquad\forall q,q'\in\mathcal{F}_t,

holds for every frame tt and every function φt\varphi_t.

Assumption 3 (L-smoothness): There exists a constant L>0L>0, independent of tt, such that

∥Dqφt(q)−Dqφt(q′)∥2≤L∥q−q′∥2,∀q,q′∈Ft,\lVert\mathfrak{D}_q\varphi_t(q)-\mathfrak{D}_q\varphi_t(q')\rVert_2\le L\lVert q-q'\rVert_2,\qquad\forall q,q'\in\mathcal{F}_t,

holds for every frame tt and every function φt\varphi_t (note that here the Riemannian gradients are compared as vectors in the ambient space Rd+4\mathbb{R}^{d+4}).

Assumption 4 (LICQ): For every frame tt and every KKT point qt⋆q_t^\star, the active constraint Jacobian A(qt⋆)A(q_t^\star) has full row rank.

Under these assumptions, we model the evolution of the framewise KKT solutions as a constrained gradient ODE, as formalized in the following theorem.

Theorem 1 (KKT flows): Suppose Assumptions 1–4 hold, and Q=Q⊤≻0Q=Q^\top\succ0. Let q0⋆=q0∈Fq_0^\star=q_0\in\mathcal{F}. For each kk, let qk+1⋆q_{k+1}^\star be a KKT point of (13) at τk+1\tau_{k+1}, with feasible configuration qk⋆q_k^\star. As δ→0\delta\to0, these KKT conditions yield

q˙⋆(τ)=Pq⋆(τ)Q(−Q−1Dqf(τ,q⋆(τ))),\dot{q}^\star(\tau)=\mathcal{P}_{q^\star(\tau)}^Q\left(-Q^{-1}\mathfrak{D}_qf(\tau,q^\star(\tau))\right),
q⋆(τ)∈F,q⋆(0)=q0.(23)q^\star(\tau)\in\mathcal{F},\qquad q^\star(0)=q_0. \tag*{(23)}

The proof is in Appendix I. Theorem 1 suggests computing the entire retargeted motion by solving a single initial-value problem. Starting from q0q_0, integrating the projected flow over the full motion interval yields a configuration trajectory whose sampled states provide the retargeted frames. Thus, humanoid motion retargeting can be performed by integrating one ODE instead of solving a separate optimization problem at every frame.

Discrete Approximation of the KKT Flow

To compute the KKT trajectory numerically, we construct discrete approximations of the KKT flow. Define the projected velocity field

V(τ,q):=PqQ(−Q−1Dqf(τ,q)).(24)V(\tau,q):=\mathcal{P}_{q}^{Q}\left(-Q^{-1}\mathfrak{D}_qf(\tau,q)\right). \tag*{(24)}

For a step size η>0\eta>0 and a time grid τk=kη\tau_k=k\eta, the backward and forward TaSP–Euler updates for (23) are, respectively,

rk+1BE=TaSP⁡rk+1,BE(−ηV(τk+1,qk+1BE)),r_{k+1}^{\mathrm{BE}}=\operatorname{TaSP}_{r_{k+1,\mathrm{BE}}}\left(-\eta V(\tau_{k+1},q_{k+1}^{\mathrm{BE}})\right),
xkBE=xk+1BE−ηV(τk+1,qk+1BE).(25)x_k^{\mathrm{BE}}=x_{k+1}^{\mathrm{BE}}-\eta V(\tau_{k+1},q_{k+1}^{\mathrm{BE}}). \tag*{(25)}

and

rk+1FE=TaSP⁡rkFE(ηV(τk,qkFE)),r_{k+1}^{\mathrm{FE}}=\operatorname{TaSP}_{r_k^{\mathrm{FE}}}\left(\eta V(\tau_k,q_k^{\mathrm{FE}})\right),
xk+1FE=xkFE+ηV(τk,qkFE).(26)x_{k+1}^{\mathrm{FE}}=x_k^{\mathrm{FE}}+\eta V(\tau_k,q_k^{\mathrm{FE}}). \tag*{(26)}

In this section, we first establish first-order convergence of the backward TaSP–Euler method (25). We then characterize the backward update by a nonlinear system on M\mathcal{M}. For a fixed reference configuration qkq_k, we show that the SQP iterations (15) converge locally and that their limit differs from a nearby feasible solution of the backward system by O(η2)\mathcal{O}(\eta^2). Building on the fact that the forward TaSP–Euler method has already been proved to converge with order one [24, 25], we propose TaSP-KKT, an algorithm for OmniRetargeting based on the update in (26). By evaluating the velocity field at the current state and advancing the configuration explicitly, TaSP-KKT avoids the nonlinear solve and associated Newton iterations required by the backward formulation, thereby reducing the computational cost per step.

Backward Euler Method and SQP Procedure

We establish the convergence order of the backward TaSP–Euler method.

Lemma 1. Suppose that Assumptions 1–4 hold, and let q:[0,T]→Fq : [0,T] \to\mathcal{F} be an exact C2C^{2} solution of (23). Then, for all sufficiently small step sizes η>0\eta> 0, the backward TaSP–Euler iterates initialized at q0BE=q(0)q_{0}^{\mathrm{BE}}=q(0) satisfy the first-order global error bound

max⁡0≤k≤nη∥qkBE−q(τk)∥≤CBE,Tη,(27)\max_{0\leq k\leq n_{\eta}}\left\|q_{k}^{\mathrm{BE}}-q(\tau_{k})\right\|\leq C_{\mathrm{BE},T}\eta, \tag*{(27)}

where nη=⌊T/η⌋n_{\eta}=\lfloor T/\eta\rfloor.

The proof is provided in Appendix II. We next express the implicit backward TaSP update in (25) as an equivalent nonlinear system.

Lemma 2. Suppose that Assumptions 1–4 hold. Let Q=QT≻0Q=Q^{\mathsf{T}}\succ0, η>0\eta>0, and qkBE∈Fq_{k}^{\mathrm{BE}}\in\mathcal{F}. Set q=qk+1BE=(r,x)∈Fq=q_{k+1}^{\mathrm{BE}}=(r,x)\in\mathcal{F}, and set I=I(q)\mathcal{I}=\mathcal{I}(q). If η∥Vr(τk+1,q)∥<1\eta\|V_{r}(\tau_{k+1},q)\|<1, then qq satisfies the backward TaSP update (25) if and only if there exists a multiplier vector πk+1=(λI,μ)\pi_{k+1}=(\lambda_{\mathcal{I}},\mu) such that

0=ΠqQΠq(q−qkBE)+ηDqfk+1(q)+ηA(q)Tπk+1,0=\Pi_{q}Q\Pi_{q}(q-q_{k}^{\mathrm{BE}})+\eta D_{q}f_{k+1}(q)+\eta A(q)^{\mathsf{T}}\pi_{k+1},
0=⟨πk+1,A(q)Πq(q−qkBE)⟩,0=\left\langle\pi_{k+1},A(q)\Pi_{q}(q-q_{k}^{\mathrm{BE}})\right\rangle,
λI≥0,q∈F.(28)\lambda_{\mathcal{I}}\geq0,\qquad q\in\mathcal{F}. \tag*{(28)}

The proof is in Appendix III.

Theorem 2 (SQP Convergence). Fix a reference configuration p=qk∈Fp=q_{k}\in\mathcal{F} and the frame data at τk+1\tau_{k+1}. Let I=I(p)\mathcal{I}=\mathcal{I}(p), c=(gI,h)c=(g_{\mathcal{I}},h), and B(q):=Dc(q)ΠqB(q):=D c(q)\Pi_{q}, with the index set I\mathcal{I} held fixed. Set δ=η\delta=\eta in (15), solve each SQP subproblem exactly, and apply the full normalization update (16). For sufficiently small η>0\eta>0, the SQP iterates converge to a limit

qk+1SQP,∞(η):=lim⁡i→∞qˉk+1,i;q_{k+1}^{\mathrm{SQP},\infty}(\eta):=\lim_{i\to\infty}\bar{q}_{k+1,i};

moreover, there exists a constant C>0C>0, independent of η\eta, such that

∥qk+1SQP,∞(η)−qk+1BE(η)∥≤Cη2.(29)\left\|q_{k+1}^{\mathrm{SQP},\infty}(\eta)-q_{k+1}^{\mathrm{BE}}(\eta)\right\|\leq C\eta^{2}. \tag*{(29)}

The proof is left in Appendix IV. Theorem 2 shows that, for a fixed reference configuration qkq_{k}, the SQP iterations converge locally to a limit within O(η2)O(\eta^{2}) of qk+1BEq_{k+1}^{\mathrm{BE}}. By Lemma 1, the backward iterates approximate the KKT flow (23) with O(η)O(\eta) global error. This connection provides a KKT-flow interpretation of the SQP limit together with a local convergence guarantee for the SQP iterations.

Forward TaSP–Euler: The TaSP-KKT Method

We first state the convergence order of the forward TaSP–Euler update (26).

Theorem 3 (First-order convergence of forward TaSP). Suppose that Assumptions 1–4 hold. Let q:[0,T]→Fq : [0,T]\to\mathcal{F} be an exact C2C^{2} solution of (23). Then the forward TaSP–Euler iterates in (26), initialized at q0FE=q(0)q_{0}^{\mathrm{FE}}=q(0), satisfy

max⁡0≤k≤nη∥qkFE−q(τk)∥2≤CFE,Tη,(30)\max_{0\leq k\leq n_{\eta}}\left\|q_{k}^{\mathrm{FE}}-q(\tau_{k})\right\|_{2}\leq C_{\mathrm{FE},T}\eta, \tag*{(30)}

for all sufficiently small $\eta>0.

Proof. See, for example, [24], [25]. □

Algorithm 1 TaSP-KKT: Forward Integration of the KKT Flow
Require: Initial state q0=(r0,x0)∈Fq_{0}=(r_{0},x_{0})\in\mathcal{F}, Q≻0Q\succ0, step size η>0\eta>0, horizon $T
Ensure: Configurations {qk}k=0nη\{q_{k}\}_{k=0}^{n_{\eta}}
1: nη←⌊T/η⌋n_{\eta}\leftarrow\lfloor T/\eta\rfloor
2: for k=0,…,nη−1k=0,\ldots,n_{\eta}-1 do
3: τk←kη\tau_{k}\leftarrow k\eta
4: vk←PqkQ(−Q−1Dqf(τk,qk))v_{k}\leftarrow P_{q_{k}}^{Q}(-Q^{-1}D_{q}f(\tau_{k},q_{k}))
5: Decompose vk=(vr,k;vx,k)v_{k}=(v_{r,k};v_{x,k})
6: sk←ηvr,ks_{k}\leftarrow\eta v_{r,k}
7: rk+1←1−∣sk∣22rk+skr_{k+1}\leftarrow\sqrt{1-|s_{k}|_{2}^{2}}r_{k}+s_{k}
8: xk+1←xk+ηvx,kx_{k+1}\leftarrow x_{k}+\eta v_{x,k}
9: qk+1←(rk+1,xk+1)q_{k+1}\leftarrow(r_{k+1},x_{k+1})
10: end for
11: return {qk}k=0nη\{q_{k}\}_{k=0}^{n_{\eta}}

Table 1.

Applying the forward TaSP–Euler method to the KKT flow yields an explicit algorithm, which we call TaSP-KKT (tangent-space parametrization for KKT flow). We summarize TaSP-KKT in Algorithm 1.

TaSP-KKT uses only the current frame data and configuration. Each integration step requires one weighted tangent-cone projection and a closed-form TaSP update, avoiding the repeated linearizations and QP solves within the framewise SQP loop. By Lemma 1 and Theorem 3, forward and backward TaSP–Euler both have first-order global accuracy under their stated assumptions. Thus, TaSP-KKT reduces the computational work per integration step while retaining the same convergence order.

Experiments and Discussions

In this section, we validate our theoretical findings via some numerical experiments.

Experimental Setup

We evaluate TaSP-KKT for retargeting human–object interaction motions to the Unitree G1 humanoid. Human demonstrations are drawn from sequences involving the large box object (largebox) in the OMOMO dataset [26], with a sampling interval of δ=1/30\delta= 1/30 s. Using the dataset’s action annotations,1 we group the selected motions into seven patterns: Lift–Move–Place, Lift–Rotate–Replace, Push–Replace, Kick–Replace, Kick–Lift–Move–Place, Push–Release–Pull–Replace, and Pull–Replace. For each pattern, we select three complete sequences from three different subjects, yielding 21 sequences in total.

We evaluate both interaction preservation and computational efficiency. The first metric is the mean interaction loss, computed by averaging the framewise loss within each sequence and then taking an equally weighted average across the three sequences in each motion pattern:

Eˉ:=13∑i=131Mi∑t=0Mi−1fi,t(qi,t),(31)\bar{E} := \frac{1}{3} \sum_{i=1}^{3} \frac{1}{M_i} \sum_{t=0}^{M_i-1} f_{i,t}(q_{i,t}), \tag*{(31)}

where Mi−1M_i - 1 is the number of frames in sequence ii, qi,tq_{i,t} is the retargeted robot configuration at frame tt, and fi,tf_{i,t} is the corresponding interaction loss. A lower Eˉ\bar{E} indicates better preservation of the reference interaction structure.

The second metric is the total retargeting solve time, Stotal:=∑i=13SiS_{\mathrm{total}} := \sum_{i=1}^{3} S_i, where SiS_i denotes the solve time for the complete sequence ii. Since sequence lengths may differ across motion patterns, we compare solve times between methods using the same sequences within each pattern.

Results and Discussion

As shown in Figure 1, TaSP-KKT requires approximately 38%–63% less computation time than OmniRetarget while yielding slightly higher mean interaction loss. The absolute differences in mean interaction loss are on the order of 10−210^{-2} for most motion patterns. These results illustrate our theoretical findings.

OMOMO largebox results

Figure 1. OMOMO largebox results: (a) mean interaction loss; (b) total solve time.

Conclusion

In this paper, we have presented an ODE perspective on interaction-preserving motion retargeting by characterizing the continuous-time limit of OmniRetarget’s framewise KKT conditions on the robot configuration manifold. This perspective supports both convergence analysis and algorithm design. By connecting the backward TaSP–Euler discretization to OmniRetarget’s SQP iterations, we establish a local convergence guarantee for the solver under suitable regularity assumptions. By discretizing the same KKT flow with a forward TaSP–Euler scheme, we derive TaSP-KKT, which achieves the same first-order convergence to the ODE solution as the backward method under the stated assumptions. Its explicit manifold update preserves the unit quaternion constraint for admissible step sizes, while replacing repeated inner SQP iterations with one weighted tangent-cone projection and a closed-form update per step. Numerical experiments demonstrate that TaSP-KKT runs approximately twice as fast as OmniRetarget while maintaining comparable retargeting quality, with a slight increase in mean interaction loss on the order of 10−210^{-2} for most motion patterns.

In future work, we will evaluate TaSP-KKT on a broader range of datasets and assess its effectiveness in downstream tasks, such as humanoid control based on reinforcement learning.

Appendix I

Proof of Theorem 1

The manifold KKT conditions [23] retain the usual primal feasibility, dual feasibility, and complementary slackness conditions. Stationarity is expressed in the tangent space using Riemannian gradients. For the fixed reference configuration qk∗q_k^*, the temporal penalty has Euclidean and Riemannian gradients

∇q[12δ∥q−qk∗∥Q2]=δ−1Q(q−qk∗),\nabla_q \left[\frac{1}{2\delta}\lVert q-q_k^*\rVert_Q^2\right] = \delta^{-1}Q(q-q_k^*),
Dq[12δ∥q−qk∗∥Q2]=δ−1ΠqQ(q−qk∗).\mathcal{D}_q \left[\frac{1}{2\delta}\lVert q-q_k^*\rVert_Q^2\right] = \delta^{-1}\Pi_qQ(q-q_k^*).

Evaluating these expressions at qk+1∗q_{k+1}^* gives the framewise KKT conditions

0=δ−1Πqk+1∗Q(qk+1∗−qk∗)+Dqf(τk+1,qk+1∗)+A(qk+1∗)⊤νk+1∗,qk+1∗∈F,λk+1∗≥0,λk+1,i∗gi(qk+1∗)=0,i=1,…,m.(32)\begin{aligned} 0 &= \delta^{-1}\Pi_{q_{k+1}^*}Q(q_{k+1}^*-q_k^*) \\ &\quad+ \mathcal{D}_q f(\tau_{k+1},q_{k+1}^*) + A(q_{k+1}^*)^\top\nu_{k+1}^*, \\ q_{k+1}^* &\in\mathcal{F}, \qquad\lambda_{k+1}^* \ge0, \\ \lambda_{k+1,i}^*g_i(q_{k+1}^*) &= 0,\qquad i=1,\ldots,m. \tag*{(32)} \end{aligned}

Here, νk+1∗\nu_{k+1}^* stacks the active inequality multipliers and the equality multipliers. Taking δ→0\delta\to0 and τk→τ\tau_k \to\tau, we have

qk∗, qk+1∗⟶q∗(τ),qk+1∗−qk∗δ⟶q˙∗(τ),q_k^*,\ q_{k+1}^* \longrightarrow q^*(\tau),\qquad\frac{q_{k+1}^*-q_k^*}{\delta} \longrightarrow\dot{q}^*(\tau),
(λk+1∗,μk+1∗)⟶(λ∗(τ),μ∗(τ)).(\lambda_{k+1}^*,\mu_{k+1}^*) \longrightarrow(\lambda^*(\tau),\mu^*(\tau)).

Therefore, we have

0=Πq∗(τ)Qq˙∗(τ)+Dqf(τ,q∗(τ))+A(q∗(τ))⊤ν∗(τ).(33)0 = \Pi_{q^*(\tau)}Q\dot{q}^*(\tau) + \mathcal{D}_q f(\tau,q^*(\tau)) + A(q^*(\tau))^\top\nu^*(\tau). \tag*{(33)}

Set n∗(τ):=A(q∗(τ))⊤ν∗(τ)∈NC(q∗(τ))n^*(\tau) := A(q^*(\tau))^\top\nu^*(\tau) \in\mathcal{N}_{\mathcal{C}}(q^*(\tau)). Since the trajectory q∗(τ)q^*(\tau) is feasible, both the forward and backward tangent directions, q˙∗(τ)\dot{q}^*(\tau) and −q˙∗(τ)-\dot{q}^*(\tau), must belong to TC(q∗(τ))\mathcal{T}_{\mathcal{C}}(q^*(\tau)). The polar-cone characterization therefore gives

⟨n∗(τ),q˙∗(τ)⟩=0.\langle n^*(\tau),\dot{q}^*(\tau)\rangle= 0.

For every w∈TC(q∗(τ))w \in\mathcal{T}_{\mathcal{C}}(q^*(\tau)), both ww and q˙∗\dot{q}^* are tangent to M\mathcal{M}. Hence (33) implies

⟨ΠqQq˙∗+Dqf,w−q˙∗⟩=⟨Qq˙∗+Dqf,w−q˙∗⟩=−⟨n∗,w−q˙∗⟩≥0.\langle\Pi_qQ\dot{q}^*+\mathcal{D}_qf,w-\dot{q}^*\rangle = \langle Q\dot{q}^*+\mathcal{D}_qf,w-\dot{q}^*\rangle = -\langle n^*,w-\dot{q}^*\rangle\ge0.

This is the first-order optimality condition for minimizing 12w⊤Qw+⟨Dqf,w⟩\frac{1}{2}w^\top Qw+\langle\mathcal{D}_qf,w\rangle over the tangent cone. Since Q≻0Q\succ0, the minimizer is unique. Thus, (33) yields

q˙∗(τ)=Pq∗(τ)Q(−Q−1Dqf(τ,q∗(τ))).(34)\dot{q}^*(\tau) = P_{q^*(\tau)}^Q\left(-Q^{-1}\mathcal{D}_qf(\tau,q^*(\tau))\right). \tag*{(34)}

The proof is completed. □\square

Proof of Lemma 1

Proof. For brevity, write qk=qkBEq_k=q_k^{\mathrm{BE}} and Vk+1:=V(τk+1,qk+1)V_{k+1}:=V(\tau_{k+1},q_{k+1}). The boundedness and Lipschitz continuity of ftf_t, gtg_t, and hth_t imply that there exist constants M>0M>0 and LV>0L_V>0, independent of τ\tau and η\eta, such that

∥V(τ,u)∥≤M,\lVert V(\tau,u)\rVert\le M,
∥V(τ,u)−V(τ,v)∥≤LV∥u−v∥.\lVert V(\tau,u)-V(\tau,v)\rVert\le L_V\lVert u-v\rVert.

The backward TaSP update (25) gives

rk=1−η2∥Vr,k+1∥2 rk+1−ηVr,k+1,r_k=\sqrt{1-\eta^2\lVert V_{r,k+1}\rVert^2}\,r_{k+1}-\eta V_{r,k+1},
xk=xk+1−ηVx,k+1.x_k=x_{k+1}-\eta V_{x,k+1}.

Consequently,

qk+1=qk+ηVk+1+Rk+1,q_{k+1}=q_k+\eta V_{k+1}+R_{k+1},

where

Rk+1=((1−1−η2∥Vr,k+1∥2)rk+10).R_{k+1}=\begin{pmatrix} \left(1-\sqrt{1-\eta^2\lVert V_{r,k+1}\rVert^2}\right)r_{k+1}\\ 0 \end{pmatrix}.

For 0<η≤(2M)−10<\eta\le(2M)^{-1}, rationalizing the square root yields

∥Rk+1∥=η2∥Vr,k+1∥21+1−η2∥Vr,k+1∥2≤M2η2.\lVert R_{k+1}\rVert=\frac{\eta^2\lVert V_{r,k+1}\rVert^2}{1+\sqrt{1-\eta^2\lVert V_{r,k+1}\rVert^2}}\le M^2\eta^2.

We introduce the frozen-field implicit sequence

qˉ0=q0,qˉk+1=qˉk+ηV(τk+1,qk+1).\bar q_0=q_0,\qquad\bar q_{k+1}=\bar q_k+\eta V(\tau_{k+1},q_{k+1}).

This iteration has the form of the classical backward Euler method, with the velocity evaluated at the backward TaSP iterate qk+1q_{k+1}. Subtracting the two recursions and summing gives qk−qˉk=∑j=1kRjq_k-\bar q_k=\sum_{j=1}^{k}R_j. Hence, for kη≤Tk\eta\le T,

∥qk−qˉk∥≤kM2η2≤M2Tη.\lVert q_k-\bar q_k\rVert\le kM^2\eta^2\le M^2T\eta.

Next, we interpolate the auxiliary sequence linearly, i.e.,

qˉ(s)=qˉj−1+(s−τj−1)Vj,s∈[τj−1,τj].\bar q(s)=\bar q_{j-1}+(s-\tau_{j-1})V_j,\qquad s\in[\tau_{j-1},\tau_j].

The velocity bound implies ∥qˉj−qˉ(s)∥≤Mη\lVert\bar q_j-\bar q(s)\rVert\le M\eta. Therefore,

∥qj−qˉ(s)∥≤∥qj−qˉj∥+∥qˉj−qˉ(s)∥≤(M2T+M)η.\lVert q_j-\bar q(s)\rVert\le\lVert q_j-\bar q_j\rVert+\lVert\bar q_j-\bar q(s)\rVert \le(M^2T+M)\eta.

Let e(s):=∥qˉ(s)−q(s)∥e(s):=\lVert\bar q(s)-q(s)\rVert. Since q˙=V(τ,q)\dot q=V(\tau,q) and ∥q˙∥≤M\lVert\dot q\rVert\le M,

∥q(τj)−q(s)∥≤Mη.\lVert q(\tau_j)-q(s)\rVert\le M\eta.

It follows that ∥qj−q(τj)∥≤e(s)+(M2T+2M)η\lVert q_j-q(\tau_j)\rVert\le e(s)+(M^2T+2M)\eta. Since q∈C2([0,T])q\in C^2([0,T]), the mean value estimate gives

∥q˙(τj)−q˙(s)∥≤ηmax⁡0≤τ≤T∥q¨(τ)∥=O(η),\lVert\dot q(\tau_j)-\dot q(s)\rVert\le\eta\max_{0\le\tau\le T}\lVert\ddot q(\tau)\rVert=O(\eta),

uniformly in jj and s∈[τj−1,τj]s\in[\tau_{j-1},\tau_j]. Thus, for s∈[τj−1,τj]s\in[\tau_{j-1},\tau_j],

∥Vj−V(s,q(s))∥≤∥V(τj,qj)−V(τj,q(τj))∥+∥q˙(τj)−q˙(s)∥≤LV∥qj−q(τj)∥+O(η)≤LVe(s)+C∗η,\begin{aligned} \lVert V_j-V(s,q(s))\rVert &\le\lVert V(\tau_j,q_j)-V(\tau_j,q(\tau_j))\rVert\\ &\quad+\lVert\dot q(\tau_j)-\dot q(s)\rVert\\ &\le L_V\lVert q_j-q(\tau_j)\rVert+O(\eta)\\ &\le L_Ve(s)+C_*\eta, \end{aligned}

where C∗>0C_*>0 is independent of η\eta, jj, and ss.

Because qˉ(0)=q(0)\bar q(0)=q(0), the integral representations of qˉ\bar q and qq give

e(τ)≤LV∫0τe(s) ds+C∗τη,0≤τ≤τnη.e(\tau)\le L_V\int_0^\tau e(s)\,\mathrm{d}s+C_*\tau\eta,\qquad0\le\tau\le\tau_{n_\eta}.

Gronwall’s inequality therefore yields

e(τ)≤C∗τeLVτη.e(\tau)\le C_*\tau e^{L_V\tau}\eta.

At the grid points,

∥qk−q(τk)∥≤∥qk−qˉk∥+∥qˉk−q(τk)∥≤(M2T+C∗TeLVT)η.\lVert q_k-q(\tau_k)\rVert\le\lVert q_k-\bar q_k\rVert+\lVert\bar q_k-q(\tau_k)\rVert \le(M^2T+C_*Te^{L_VT})\eta.

Thus (27) holds with

CBE,T:=M2T+C∗TeLVT,C_{\mathrm{BE},T}:=M^2T+C_*Te^{L_VT},

which is independent of η\eta. □

Proof of Lemma 2

Proof. Set Vk+1:=V(τk+1,q)V_{k+1}:=V(\tau_{k+1},q), dk+1=Πq(q−qk+1BE)d_{k+1}=\Pi_q(q-q_{k+1}^{\mathrm{BE}}). TaSP inverts tangent projection, so (25) ⟺\Longleftrightarrow

dk+1=ηVk+1.d_{k+1}=\eta V_{k+1}.

Applying (21) with u=−Q−1Dqfk+1(q)u=-Q^{-1}D_qf_{k+1}(q) and w=dk+1/ηw=d_{k+1}/\eta, there exist πk+1,I≥0\pi_{k+1,I}\ge0 such that

−Dqfk+1(q)−η−1ΠqQdk+1=A(q)⊤πk+1∈NC(qt),-D_qf_{k+1}(q)-\eta^{-1}\Pi_qQd_{k+1}=A(q)^\top\pi_{k+1}\in NC(q_t),
⟨Q(u−w),w⟩=η−1⟨πk+1,I,A(q)dk+1⟩=0.\langle Q(u-w),w\rangle=\eta^{-1}\langle\pi_{k+1,I},A(q)d_{k+1}\rangle=0.

Together with q∈Fq\in\mathcal{F}, these relations complete the proof. □

Proof of Theorem 2

Proof. Fix the reference configuration p=qkp=q_k. Set Ji−1=Jk,i−1eJ_{i-1}=J_{k,i-1}^{e} and qˉi−1=qˉk,i−1\bar q_{i-1}=\bar q_{k,i-1}. With the trust-region constraint omitted, the SQP increment Δq\Delta_q at iteration i−1i-1 solves

min⁡Δq∈TqMη∥e(qˉi−1)+Ji−1Δq∥2+12∥qˉi−1+Δq−p∥Q2s.t.g(qˉi−1)+Jt,i−1gΔq≤0,h(qˉi−1)+Jt,i−1hΔq=0.(35)\begin{aligned} \min_{\Delta_q\in T_q\mathcal{M}}\quad& \eta\lVert e(\bar q_{i-1})+J_{i-1}\Delta_q\rVert^2 +\frac{1}{2}\lVert\bar q_{i-1}+\Delta_q-p\rVert_Q^2\\ \text{s.t.}\quad& g(\bar q_{i-1})+J_{t,i-1}^{g}\Delta_q\le0,\\ &h(\bar q_{i-1})+J_{t,i-1}^{h}\Delta_q=0. \tag*{(35)} \end{aligned}

Define Hi−1:=Q+2ηJi−1⊤Ji−1H_{i-1}:=Q+2\eta J_{i-1}^{\top}J_{i-1}, Fi−1:=2Ji−1⊤e(qˉi−1)F_{i-1}:=2J_{i-1}^{\top}e(\bar q_{i-1}). The KKT conditions in M\mathcal{M} are

0=Πqˉi−1(Hi−1Δq+Q(qˉi−1−p)+ηFi−1)+Πqˉt−1(Jt,i−1g)⊤λ+Πqˉt−1(Jt,i−1h)⊤μ,Δq∈TqM,g(qˉi−1)+Jt,i−1gΔq≤0,h(qˉi−1)+Jt,i−1hΔq=0,λ∈R+m,μ∈Rℓ,λj[gj(qˉi−1)+Jt,i−1gjΔq]=0,j=1,…,m.(36)\begin{aligned} 0={}&\Pi_{\bar q_{i-1}}\left(H_{i-1}\Delta_q+Q(\bar q_{i-1}-p)+\eta F_{i-1}\right)\\ &+\Pi_{\bar q_{t-1}}\left(J_{t,i-1}^{g}\right)^\top\lambda +\Pi_{\bar q_{t-1}}\left(J_{t,i-1}^{h}\right)^\top\mu,\\ &\Delta_q\in T_q\mathcal{M},\\ &g(\bar q_{i-1})+J_{t,i-1}^{g}\Delta_q\le0,\\ &h(\bar q_{i-1})+J_{t,i-1}^{h}\Delta_q=0,\\ &\lambda\in\mathbb{R}_+^m,\qquad\mu\in\mathbb{R}^{\ell},\\ &\lambda_j\left[g_j(\bar q_{i-1})+J_{t,i-1}^{g_j}\Delta_q\right]=0,\qquad j=1,\ldots,m. \tag*{(36)} \end{aligned}

Therefore, the SQP iteration is the composition

qˉi−1↦Δq↦qˉi:=U(qˉi−1):=(rˉi−1+Δr∥rˉi−1+Δr∥,xˉi−1+Δx).\bar q_{i-1}\mapsto\Delta_q\mapsto\bar q_i:=\mathcal{U}(\bar q_{i-1}):=\left(\frac{\bar r_{i-1}+\Delta r}{\lVert\bar r_{i-1}+\Delta r\rVert},\bar x_{i-1}+\Delta x\right).

Now we consider the trust region bound ϵ\epsilon, we let ϵ=Rδ\epsilon= R\delta and define the neighborhood

Dη:={q∈M:∥q−p∥≤ϵ}.\mathcal{D}_{\eta} := \{q \in\mathcal{M} : \lVert q - p \rVert\le\epsilon\}.

We show that, for sufficiently small η\eta, the map Uη\mathcal{U}_{\eta} is a local contraction on the Dη\mathcal{D}_{\eta}, i.e.,

∥U(q)−U(v)∥≤β∥q−v∥,U(Dη)⊆Dη.β<1(37)\lVert\mathcal{U}(q) - \mathcal{U}(v) \rVert\le\beta\lVert q - v \rVert,\quad\mathcal{U}(\mathcal{D}_{\eta}) \subseteq\mathcal{D}_{\eta}.\quad\beta< 1 \tag*{(37)}

We omit the proof due to space constraints. By the contraction property of U\mathcal{U} on Dη\mathcal{D}_{\eta}, the iteration converges to a fixed point qˉ∞\bar{q}_{\infty} at which Δq=0\Delta q = 0. Hence, the KKT condition become

0=Πqˉ∞Q(qˉ∞−p)+Πqˉ∞ηFqˉ∞+Πqˉt−1(Jt,i−1g)⊤λ+Πqˉt−1(Jt,i−1h)⊤μ.0 = \Pi_{\bar{q}_{\infty}}Q(\bar{q}_{\infty} - p) + \Pi_{\bar{q}_{\infty}}\eta F_{\bar{q}_{\infty}} + \Pi_{\bar{q}_{t-1}}(J^{g}_{t,i-1})^{\top}\lambda+ \Pi_{\bar{q}_{t-1}}(J^{h}_{t,i-1})^{\top}\mu.

Given that

Πqˉ∞Fqˉ∞=Πqˉ∞∇qfk+1(qˉ∞)=Dfk+1(qˉ∞),\Pi_{\bar{q}_{\infty}}F_{\bar{q}_{\infty}} = \Pi_{\bar{q}_{\infty}}\nabla_{q}f_{k+1}(\bar{q}_{\infty}) = \mathcal{D}f_{k+1}(\bar{q}_{\infty}),
Πqˉt−1(Jt,i−1g)=Dgk+1(qˉ∞),Πqˉt−1(Jt,i−1h)=Dhk+1(qˉ∞),\Pi_{\bar{q}_{t-1}}(J^{g}_{t,i-1}) = \mathcal{D}g_{k+1}(\bar{q}_{\infty}),\quad\Pi_{\bar{q}_{t-1}}(J^{h}_{t,i-1}) = \mathcal{D}h_{k+1}(\bar{q}_{\infty}),

therefore, with muliter π∞=(λI∞,μ∞)\pi_{\infty} = (\lambda_{\mathcal{I}_{\infty}},\mu_{\infty}), we have

0=Πqˉ∞Q(qˉ∞−p)+ηDqfk+1(qˉ∞)+A(qˉ∞)⊤π∞,0 = \Pi_{\bar{q}_{\infty}}Q(\bar{q}_{\infty} - p) + \eta\mathcal{D}q f_{k+1}(\bar{q}_{\infty}) + A(\bar{q}_{\infty})^{\top}\pi_{\infty},
λI∞≥0,qˉ∞∈F.\lambda_{\mathcal{I}_{\infty}} \ge0,\quad\bar{q}_{\infty} \in\mathcal{F}.

Compared with (28), this system use Πqˉ∞Q\Pi_{\bar{q}_{\infty}}Q instead of ΠqQΠq\Pi_{q}Q\Pi_{q}. For a backward solution u=qk+1BEu = q^{\mathrm{BE}}_{k+1} with u−p=O(η)u - p = O(\eta), its stationarity residual in the above system is

bη=ΠuQ(I−Πu)(u−p).b_{\eta} = \Pi_{u}Q(I - \Pi_{u})(u - p).

The unit-quaternion identity gives

∥bη∥≤∥Q∥2∥ru−rp∥2=O(η2).\lVert b_{\eta}\rVert\le\frac{\lVert Q\rVert}{2}\lVert r_{u} - r_{p}\rVert^{2} = O(\eta^{2}).

The local KKT inverse estimate therefore yields

∥qˉ∞−qk+1BE∥≤C∥bη∥=O(η2).\lVert\bar{q}_{\infty} - q^{\mathrm{BE}}_{k+1}\rVert\le C\lVert b_{\eta}\rVert= O(\eta^{2}).

References

  1. [1]Z. Gu, et al., “Humanoid locomotion and manipulation: Current progress and challenges in control, planning, and learning,” IEEE/ASME Trans. Mechatronics, vol. 31, no. 2, pp. 2300–2330, 2026.arxiv.org/abs/2501.02116
  2. [2]Y. Gu, et al., “Evolution of humanoid locomotion control,” Sci. Robot., vol. 11, no. 117, p. eead3973, 2026.DOI
  3. [3]M. Gleicher, “Retargetting motion to new characters,” in Proc. 25th Annu. Conf. Comput. Graph. Interactive Techn., 1998, pp. 33–42.DOI
  4. [4]K.-J. Choi and H.-S. Ko, “Online motion retargetting,” J. Vis. Comput. Animat., vol. 11, no. 5, pp. 223–235, 2000.doi.org/10.1002/1099-1778(200012)11:5<223::aid-vis236>3.0.co;2-5
  5. [5]J. P. Araújo, Y. Ze, P. Xu, J. Wu, and C. K. Liu, “Retargeting matters: General motion retargeting for humanoid motion tracking,” in Proc. IEEE Int. Conf. Robot. Automat., 2026.arxiv.org/abs/2510.02252
  6. [6]R. Villegas, D. Ceylan, A. Hertzmаnn, J. Yang, and J. Saito, “Contact-aware retargeting of skinned motion,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2021, pp. 9720–9729.arxiv.org/abs/2109.07431
  7. [7]J. Zhang, Z. Tu, J. Weng, J. Yuan, and B. Du, “A modular neural motion retargeting system decoupling skeleton and shape perception,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 46, no. 10, pp. 6889–6904, 2024.DOI
  8. [8]L. Yang, et al., “OmniRetarget: Interaction-preserving data generation for humanoid whole-body loco-manipulation and scene interaction,” in Proc. IEEE Int. Conf. Robot. Automat., 2026.arxiv.org/abs/2509.26633
  9. [9]U. Helmke and J. B. Moore, Optimization and Dynamical Systems, ser. Communications and Control Engineering. Springer London, 1994.
  10. [10]J. Schropp and I. Singer, “A dynamical systems approach to constrained minimization,” Numer. Funct. Anal. Optim., vol. 21, no. 3–4, pp. 537–551, 2000.DOI
  11. [11]S. Fiori, “Quasi-geodesic neural learning algorithms over the orthogonal group: A tutorial,” J. Mach. Learn. Res., vol. 6, no. 26, pp. 743–781, 2005.
  12. [12]E. Haber, L. Rutthotto, E. Holtham, and S.-H. Jun, “Learning across scales—multiscale methods for convolution neural networks,” in Proc. 32nd AAAI Conf. Artif. Intell., 2018.DOI
  13. [13]C. Lu, Y. Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,” in Adv. Neural Inf. Process. Syst., vol. 35, 2022.arxiv.org/abs/2206.00927
  14. [14]H. Klein, N. Jaquier, A. Meixner, and T. Asfour, “A Riemannian take on human motion analysis and retargeting,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2022, pp. 5210–5217.arxiv.org/abs/2208.01372
  15. [15]———, “On the design of region-avoiding metrics for collision-safe motion generation on Riemannian manifolds,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2023, pp. 2346–2353.arxiv.org/abs/2307.15440
  16. [16]J. Li, X. Song, S. Wang, H. Huang, and Y. Zhao, “A unified conditional flow for motion generation, editing, and intra-structural retargeting,” arXiv preprint arXiv:2604.13427, 2026.DOI
  17. [17]W. Kim, T. Li, and S. Ha, “MoReFlow: Motion retargeting learning through unsupervised flow matching,” arXiv preprint arXiv:2509.25600, 2025.arxiv.org/abs/2509.25600
  18. [18]E. Hairer and C. Lubich, “Numerical solution of ordinary differential equations,” in The Princeton Companion to Applied Mathematics. Princeton University Press, 2015, pp. 293–305.
  19. [19]E. Hairer, C. Lubich, and G. Wanner, Geometric Numerical Integration: Structure-Preserving Algorithms for Ordinary Differential Equations, 2nd ed., ser. Springer Series in Computational Mathematics. Springer, 2004, vol. 31.
  20. [20]E. Hairer, “Geometric integration of ordinary differential equations on manifolds,” BIT Numer. Math., vol. 41, no. 5, pp. 996–1007, 2001.DOI
  21. [21]H. Si and K. Gärtner, “Meshwise piecewise linear complexes by constrained Delaunay tetrahedralizations,” in Proc. 14th Int. Meshing Roundtable. Springer, 2005, pp. 147–163.DOI
  22. [22]O. Sorkine, D. Cohen-Or, Y. Lipman, M. Alexa, C. Rössl, and H.-P. Seidel, “Laplacian surface editing,” in Proc. Eurographics/ACM SIGGRAPH Symp. Geometry Process., 2004.
  23. [23]R. Bergmann and R. Herzog, “Intrinsic formulation of KKT conditions and constraint qualifications on smooth manifolds,” SIAM J. Optim., vol. 29, no. 4, pp. 2423–2444, 2019.DOI
  24. [24]E. Hairer and G. Wanner, Solving Ordinary Differential Equations II: Stiff and Differential-Algebraic Problems, 2nd ed., ser. Springer Series in Computational Mathematics. Springer, 1996, vol. 14.
  25. [25]V. Solo and Z. Wang, “Convergence of a tangent space numerical scheme for a stochastic differential equation on a sphere,” in Proc. 60th IEEE Conf. Decis. Control, 2021, pp. 5814–5819.DOI
  26. [26]J. Li, J. Wu, and C. K. Liu, “Object motion guided human motion synthesis,” ACM Trans. Graph., 2023.arxiv.org/abs/2309.16237

Paper details

vlnc:W_3w8fugck
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents