betaFeedback

WorksComputer Science

Kinematic-Tactile Aligned Dexterous Retargeting with Contact Wrench Space Alignment

Abstract

Human demonstrations provide a natural and scalable source of data for learning dexterous manipulation policies. However, existing retargeting methods primarily transfer human kinematics, while the tactile information underlying physical interaction is rarely preserved across embodiments. To address this limitation, we present a hierarchical three-stage framework for kinematic-tactile aligned dexterous retargeting. We use object-centric contact wrench space as a physical representation for tactile alignment between human and robot hands. The framework first aligns hand kinematics and contact geometry, and then constructs a human wrench support profile from tactile forces and contact geometry. Robot fingertip force targets are subsequently optimized to cover the human wrench support and converted into joint position corrections, yielding directly replayable robot hand trajectories. We evaluate the framework in simulation and real robot experiments across three pouring conditions. In the real robot experiments, our method achieves a 96.7% replay success rate and higher wrench transfer fidelity than representative retargeting baselines. The framework enables physically grounded transfer of kinematic and tactile information from human demonstrations to robot hands, providing a path toward scalable learning of contact-rich dexterous manipulation.

Introduction

Dexterous manipulation has long been challenging due to the complex kinematics of dexterous hands [1]. Human demonstrations provide a natural and scalable source of data for learning dexterous manipulation policies, motivating growing efforts to transfer human manipulation skills to anthropomorphic robot hands [2]. Retargeting plays a critical role in converting human data into robot hand trajectories and has been widely used in both teleoperation and robot-free data collection [3], [4]. However, existing retargeting methods primarily focus on aligning the kinematics of human and robot

Manuscript received October 15, 2026; revised Month xx, xxxx; accepted Month x, xxxx. This work was supported by the National Natural Science Foundation of China under Grants W2441018 and 62573321. (Corresponding author: Shuo Jiang.)

Xinyuan He, Ziheng Xu, Yueyuan Chen, Yufan Ye, Yuanshuo Tan, Huiming Pan, and Jianjun Meng are with the State Key Laboratory of Mechanical System and Vibration, School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China (e-mail: {hexinyuan, sj-txuz0079, yueyuanchen, yuefan18757557508, tanyuanshuo, panhuiming, mengjianjunxs008}@sjtu.edu.cn). Shuo Jiang is with the College of Electronics and Information Engineering, Tongji University, Shanghai 201804, China, and is also with the International Collaboration for Creative Research Teams, Shanghai Jiao Tong University, Shanghai 200240, China (e-mail: [email protected]). Peter B. Shull is with the School of Artificial Intelligence, The Chinese University of Hong Kong, Shenzhen, Shenzhen 518172, China (e-mail: [email protected]).

hands, while largely overlooking tactile alignment, which plays an important role in reliable and physically consistent manipulation [5]. Therefore, effective dexterous retargeting should consider both kinematic and tactile alignment to faithfully transfer human manipulation skills across embodiments.

Recent studies have made encouraging progress in functional retargeting, yet the explicit use of human tactile information remains limited. Conventional kinematic methods align human and robot hand kinematics through predefined task-space correspondences, such as fingertip positions and relationships between fingers [6]. More recent interaction-aware methods further preserve contact regions or hand–object interaction structures, improving the consistency of transferred contacts [7], [8]. However, these objectives remain geometric and provide no direct constraint on contact forces. Modern approaches increasingly employ reinforcement learning to refine retargeted motions through interaction with simulated environments [9]–[11]. Although these methods can achieve strong performance in simulation, they primarily recover executable motions through policy optimization and task-specific objectives rather than through explicit tactile alignment across embodiments. Therefore, existing retargeting methods still lack an explicit physical representation to align both kinematic and tactile information for faithful skill transfer across embodiments.

In this letter, we present a hierarchical three-stage retargeting framework that jointly aligns kinematic and tactile information across human and robot embodiments (Figure 1).

Figure 1. Overview of the hierarchical three-stage retargeting framework. (a) Human-to-robot kinematic alignment transfers reconstructed human motion to the robot embodiment. (b) Physics-aware contact alignment refines the hand–object configuration. (c) Object-centric wrench space alignment optimizes robot normal force targets to cover the human wrench support profile along the demonstrated manipulation trajectory and converts them into joint position corrections.

Our key insight is to use object-centric contact wrench space as a shared physical representation for tactile alignment across different hand embodiments. This learning-free formulation enables stable and physically consistent retargeting in real-world manipulation with varying payloads and object deformation.

Specifically, our framework progressively aligns hand kinematics, contact geometry, and object-centric wrench support. First, the kinematic alignment stage matches corresponding human and robot task-space vectors frame by frame to generate an initial robot hand trajectory. Second, the contact alignment stage uses tactile-activated contact regions to refine the human and robot hand–object configurations at an anchor frame in simulation, and propagates the resulting contact residuals throughout the trajectories. Finally, the wrench space alignment stage constructs the human object-centric wrench support profile from contact geometry and tactile forces, and optimizes robot fingertip force targets to cover this profile. These force targets are then converted into joint position corrections, yielding the final directly replayable robot hand trajectory.

We evaluate our approach in both simulation and physical experiments on a paper cup pouring task under three payload conditions with different materials and masses. The results demonstrate that our framework produces directly replayable robot hand trajectories while preserving human-derived wrench support characteristics across payload conditions. In real-world trajectory replay, our method achieves a 96.7% success rate over 60 trials, outperforming both baselines. Our method also achieves higher wrench transfer fidelity than the baselines in both the global and task-relevant wrench spaces.

In summary, our main contributions are:

  1. We introduce contact wrench space as a physical representation for tactile alignment in dexterous retargeting.

  2. We develop a hierarchical three-stage retargeting framework that progressively aligns kinematic, contact, and wrench information from human demonstrations and produces directly replayable robot hand trajectories.

  3. Extensive real robot experiments show that our method achieves higher replay success and wrench transfer fidelity than representative retargeting baselines. The framework also produces replayable robot demonstrations with human-derived wrench support, providing a practical data source for future contact-rich manipulation learning.

Data Collection System

Demonstrations for dexterous manipulation are acquired through robot teleoperation, human video, or motion capture. Teleoperation directly yields trajectories in the robot’s configuration space [6], whereas video- and mocap-based pipelines capture natural human hand–object motion [2, 4]. Existing human hand–object datasets provide detailed hand and object kinematics and, in some cases, geometric contact annotations, but generally lack direct measurements of time-varying contact forces [12, 13]. DexForce extracts force-informed actions from kinesthetic demonstrations; AET-TacNet infers tactile signals from EMG and hand kinematics; and TacDix provides synchronized tactile, kinematic, and object-state data [5, 14, 15]. Our acquisition pipeline instead combines synchronized hand–object kinematics with distributed normal force measurements to construct a time-varying human wrench support reference for retargeting.

Human-to-Robot Dexterous Hand Retargeting

Kinematics-based hand retargeting aligns joint configurations, fingertip positions, or relationships between fingers [3, 16, 17]. Because these objectives are hand-centric, small alignment errors can alter or eliminate intended contacts [7, 8]. Interaction-aware methods transfer contact areas or preserve hand–object interaction graphs, but their alignment objectives remain primarily geometric [7, 8]. Contact geometry alone does not determine the contact force distribution or resulting object wrench [5, 18]. Simulation- and policy-based approaches recover executable interactions from human demonstrations: ManipTrans and DexH2R learn residual corrections to transferred trajectories; DexMachina tracks object motion through curriculum-based virtual control; SPIDER refines kinematic references through physics-based sampling and contact guidance [9, 10, 11, 19]. These approaches improve the physical executability of retargeted motions through simulation-based refinement or policy learning. In contrast, our hierarchical framework uses demonstration-derived wrench support as an explicit retargeting objective alongside kinematic and contact geometry alignment, yielding directly replayable joint trajectories without policy learning.

Wrench Space Representations for Dexterous Manipulation

Wrench space representations characterize the external forces and torques that a set of contacts can support and are used to evaluate grasp stability and synthesize force closure grasps [20, 21, 22]. Differentiable force-closure objectives enable diverse grasp synthesis across arbitrary hand structures [23], and subsequent work accelerates this formulation for large-scale dataset generation [24]. Differentiable grasp-wrench boundary objectives further support task-oriented pose generation [25], while BOdex and Dexonomy incorporate optimization of contact force feasibility or force-based validation into scalable grasp synthesis [26, 27]. These methods synthesize or validate static hand poses under modeled contact constraints rather than transfer time-varying interaction mechanics from human demonstrations. CHORD uses object-centric human–robot contact wrench consistency to guide reinforcement learning policies [18], enabling wrench-level comparison across embodiments while obtaining robot motions through policy optimization. In our method, wrench space provides a framewise representation for physical alignment: a one-sided coverage objective optimizes robot normal force targets against the human wrench support profile, and these targets are then converted into joint position corrections.

Methods

Given synchronized trajectories of markers attached to the hand and object, together with tactile measurements, the proposed framework retargets a human manipulation demonstration to a dexterous robot hand while transferring its object-centric wrench support. Following MoCap-to-MANO reconstruction, the framework proceeds through three hierarchical stages: human-to-robot kinematic alignment, physics-aware contact alignment, and object-centric wrench space alignment. The final stage converts optimized robot normal force targets into joint position corrections, yielding a directly replayable robot joint position trajectory (Figure 2).

Figure 2. Overview of the proposed hierarchical object-centric retargeting framework. The human demonstration is recorded using motion capture and a 19-channel tactile glove, yielding synchronized hand and object kinematics and tactile normal force measurements. MoCap-to-MANO reconstruction provides the human hand sequence. (1) Human-to-robot kinematic alignment produces an initial robot trajectory. (2) Physics-aware contact alignment locally refines the human and robot contact configurations in simulation and propagates the contact residual. (3) Object-centric wrench space alignment optimizes robot normal force targets to cover the human wrench support profile, converts them into joint position corrections, and produces a replayable joint position trajectory.

MoCap-to-MANO Reconstruction

Synchronized trajectories of markers attached to the hand and object, together with tactile measurements, provide the kinematic–tactile demonstration. We represent the human hand using the MANO parametric model [28]. For each subject, we run HaMeR [29] on multiple RGB images and take the elementwise mean of its estimates of the MANO shape vector as β^\hat{\boldsymbol{\beta}}. This vector is fixed throughout marker-based sequence fitting. Each physical marker is manually associated with a MANO joint and a joint-local offset. The predicted marker position is

y^t,i=pt,k(i)+Rt,k(i)Δpi.(1)\hat{\mathbf{y}}_{t,i} = \mathbf{p}_{t,k(i)} + \mathbf{R}_{t,k(i)}\Delta\mathbf{p}_i . \tag*{(1)}

where pt,k(i)\mathbf{p}_{t,k(i)} and Rt,k(i)\mathbf{R}_{t,k(i)} are the position and orientation of MANO joint k(i)k(i), respectively, and Δpi\Delta\mathbf{p}_i is its fixed joint-local marker offset. Let ξt\boldsymbol{\xi}_t collect the MANO pose and global wrist parameters. We reconstruct the MANO sequence by solving

ξ1:T∗=arg⁡min⁡ξ1:T[λmLmarker+λpLpose+λvLvelocity+λaLacceleration].(2)\boldsymbol{\xi}^{*}_{1:T} = \arg\min_{\boldsymbol{\xi}_{1:T}} \left[\lambda_m\mathcal{L}_{\mathrm{marker}} + \lambda_p\mathcal{L}_{\mathrm{pose}} + \lambda_v\mathcal{L}_{\mathrm{velocity}} + \lambda_a\mathcal{L}_{\mathrm{acceleration}}\right]. \tag*{(2)}

Here, Lmarker\mathcal{L}_{\mathrm{marker}} is the confidence-weighted Huber loss between predicted and measured markers, Lpose\mathcal{L}_{\mathrm{pose}} penalizes anatomically implausible finger articulations, and Lvelocity\mathcal{L}_{\mathrm{velocity}} and Lacceleration\mathcal{L}_{\mathrm{acceleration}} penalize first- and second-order temporal differences, respectively. The reconstructed MANO trajectory {ξt∗}t=1T\{\boldsymbol{\xi}^{*}_t\}_{t=1}^{T} is then used for kinematic retargeting and human contact reconstruction.

Human-to-Robot Kinematic Alignment

We obtain a kinematically aligned robot hand trajectory by minimizing discrepancies between corresponding human and robot task vectors:

qtkin=arg⁡min⁡q∑i=1N∥αvti−ϕi(q)∥2+λs∥q−qt−1kin∥2\mathbf{q}^{\mathrm{kin}}_t = \arg\min_{\mathbf{q}} \sum_{i=1}^{N}\left\|\alpha\mathbf{v}^{i}_t-\boldsymbol{\phi}_i(\mathbf{q})\right\|^2+\lambda_s\left\|\mathbf{q}-\mathbf{q}^{\mathrm{kin}}_{t-1}\right\|^2
s.t.qℓ≤q≤qu.(3)\text{s.t.}\qquad\mathbf{q}_{\ell}\leq\mathbf{q}\leq\mathbf{q}_{u}. \tag*{(3)}

Here, q\mathbf{q} is the robot joint position vector being optimized, NN is the number of task vector correspondences, and vti\mathbf{v}^{i}_t and ϕi(q)\boldsymbol{\phi}_i(\mathbf{q}) denote the corresponding human and robot task vectors, respectively. The task vector set comprises wrist-to-fingertip and inter-fingertip vectors. The scale factor α\alpha compensates for the difference in hand size between the human and robot, while λs\lambda_s promotes temporal smoothness across successive frames. Joint feasibility is enforced by the bounds qℓ\mathbf{q}_{\ell} and qu\mathbf{q}_{u}. At each frame, the bound-constrained problem in Eq. (3) is solved using the SLSQP method. Solving Eq. (3) sequentially over all frames yields the kinematically aligned trajectory {qtkin}t=1T\{\mathbf{q}^{\mathrm{kin}}_t\}_{t=1}^{T}, which initializes the subsequent physics-aware contact alignment.

Physics-Aware Contact Alignment

Contact consistency is essential for faithfully transferring the mechanics of hand–object interaction from the human demonstration to the robot hand. If the intended hand–object contacts are not properly established, the wrench-guided joint position corrections are diverted toward closing contact gaps rather than reproducing the desired contact wrench, thereby reducing transfer fidelity. We therefore enforce contact consistency at an anchor frame and broadcast the resulting contact residuals across the full trajectories.

With the object pose held constant, we optimize the wrist and finger poses of both the MANO and Sharpa Wave hands in MuJoCo. For embodiment X∈{H,R}X \in\{H,R\}, where H and R denote the human and robot hands, respectively, let piX\mathbf{p}_i^X denote the position of contact site ii and pio,X\mathbf{p}_i^{o,X} its corresponding nearest point on the object. The site-level virtual force and its induced generalized force are

fiv,X=kf(pio,X−piX),τiv,X=(JiX)Tfiv,X(4)\mathbf{f}_i^{v,X}=k_f\left(\mathbf{p}_i^{o,X}-\mathbf{p}_i^X\right), \qquad\boldsymbol{\tau}_i^{v,X}=\left(\mathbf{J}_i^X\right)^\mathsf{T}\mathbf{f}_i^{v,X} \tag*{(4)}

where kfk_f is the virtual force gain and JiX\mathbf{J}_i^X is the contact Jacobian. The position difference, virtual force, and translational rows of the Jacobian are expressed in the same coordinate frame. For stage s∈{p,f}s \in\{p,f\}, where p and f denote the palm and finger stages, respectively, let CpC_p and CfC_f denote the corresponding sets of tactile-activated patches. These region-specific sets are disjoint and have different cardinalities. The total generalized force applied in stage ss is

τsv,X=∑i∈Csτiv,X,s∈{p,f}.(5)\boldsymbol{\tau}_s^{v,X}=\sum_{i\in C_s}\boldsymbol{\tau}_i^{v,X}, \qquad s\in\{p,f\}. \tag*{(5)}

Only the activated patches in CsC_s contribute to the corresponding stage.

Starting from open-hand poses, we first activate the palm patches and apply their virtual forces for 200 simulation steps to optimize only the wrist poses. We then interpolate the fingers toward their respective reference poses—reconstructed for MANO and kinematically aligned for the robot. Finally, we activate the finger patches and apply their virtual forces for another 200 simulation steps to optimize only the finger joints. A 1 mm MuJoCo contact margin is applied to prevent hand–object penetration.

We select the first trajectory frame as the anchor, ta=1t_a=1. The refinement described above yields the contact-aligned MANO and robot states ξtac\boldsymbol{\xi}_{t_a}^{c} and qtac\mathbf{q}_{t_a}^{c} at this frame. Their differences from the reconstructed MANO state ξta∗\boldsymbol{\xi}_{t_a}^{*} and the kinematically aligned robot state qtakin\mathbf{q}_{t_a}^{\mathrm{kin}} define fixed contact residuals, which are broadcast across the full trajectories:

Δξc=ξtac−ξta∗,Δqc=qtac−qtakin,\Delta\boldsymbol{\xi}^{c}=\boldsymbol{\xi}_{t_a}^{c}-\boldsymbol{\xi}_{t_a}^{*}, \qquad\Delta\mathbf{q}^{c}=\mathbf{q}_{t_a}^{c}-\mathbf{q}_{t_a}^{\mathrm{kin}},
ξtc=ξt∗+Δξc,qtc=qtkin+Δqc,t=1,…,T.(6)\boldsymbol{\xi}_t^{c}=\boldsymbol{\xi}_t^{*}+\Delta\boldsymbol{\xi}^{c}, \qquad\mathbf{q}_t^{c}=\mathbf{q}_t^{\mathrm{kin}}+\Delta\mathbf{q}^{c}, \qquad t=1,\ldots,T. \tag*{(6)}

The resulting contact-aligned MANO and robot trajectories {ξtc}t=1T\{\boldsymbol{\xi}_t^{c}\}_{t=1}^{T} and {qtc}t=1T\{\mathbf{q}_t^{c}\}_{t=1}^{T} provide the human and robot inputs to the subsequent object-centric wrench space alignment.

Object-Centric Wrench Space Alignment

Given the contact-aligned MANO and robot geometries, we optimize robot normal force targets to cover the human wrench support profile while holding contact geometry fixed. In this subsection, contact positions and basis vectors are expressed in the object frame unless otherwise stated, and moments are taken about the object center of mass. We omit the frame index tt and embodiment superscript X∈{H,R}X \in\{H,R\} during the framewise derivation.

For a hand with MM contacts, let pi∈R3\mathbf{p}_i \in\mathbb{R}^3 denote the position of contact ii and Qi=[ni,di,ei]Q_i=[\mathbf{n}_i,\mathbf{d}_i,\mathbf{e}_i] its orthonormal basis, where ni\mathbf{n}_i is the inward normal and di,ei\mathbf{d}_i,\mathbf{e}_i are unit tangent vectors satisfying ni=di×ei\mathbf{n}_i=\mathbf{d}_i\times\mathbf{e}_i. For a fixed normal force fn,i≥0f_{n,i}\geq0, the Coulomb friction model gives the admissible force set in local contact coordinates:

Fi(fn,i)={fi=[fn,ifd,ife,i]∈R3 | fd,i2+fe,i2≤μi2fn,i2}.(7)\mathcal{F}_i(f_{n,i})=\left\{\mathbf{f}_i=\begin{bmatrix}f_{n,i}\\f_{d,i}\\f_{e,i}\end{bmatrix}\in\mathbb{R}^3\ \middle|\ f_{d,i}^2+f_{e,i}^2\leq\mu_i^2f_{n,i}^2\right\}. \tag*{(7)}

where μi\mu_i is the friction coefficient; we abbreviate this set as Fi\mathcal{F}_i. With moment arm ri=pi−cO\mathbf{r}_i=\mathbf{p}_i-\mathbf{c}^{O} relative to the object center of mass cO\mathbf{c}^{O}, the normalized contact wrench map is

G~i=[QiρO−1[ri]×Qi]∈R6×3.(8)\widetilde{\mathbf{G}}_i=\begin{bmatrix}Q_i\\ \rho_O^{-1}[\mathbf{r}_i]_\times Q_i\end{bmatrix}\in\mathbb{R}^{6\times3}. \tag*{(8)}

Here, [ri]×[\mathbf{r}_i]_\times is the cross product matrix, and the object characteristic length ρO>0\rho_O>0 normalizes torque relative to force. The map converts local force components into a wrench expressed in the object frame, giving w~i=G~ifi\widetilde{\mathbf{w}}_i=\widetilde{\mathbf{G}}_i\mathbf{f}_i. For fixed normal forces and independent contact force constraints, the normalized contact wrench space (CWS) is

W~C=⨁i=1MW~i,(9)\widetilde{\mathcal{W}}_C=\bigoplus_{i=1}^{M}\widetilde{\mathcal{W}}_i, \tag*{(9)}

where ⊕\oplus denotes the Minkowski sum, collecting all sums of admissible individual contact wrenches.

For a nonempty compact set A\mathcal{A}, its support function is hA(u)=max⁡a∈AuTah_{\mathcal{A}}(\mathbf{u})=\max_{\mathbf{a}\in\mathcal{A}}\mathbf{u}^{\mathsf{T}}\mathbf{a}. We use the support function identities hLA(u)=hA(LTu)h_{\mathbf{L}\mathcal{A}}(\mathbf{u})=h_{\mathcal{A}}(\mathbf{L}^{\mathsf{T}}\mathbf{u}) for a linear map L\mathbf{L}, and h⨁iAi(u)=∑ihAi(u)h_{\bigoplus_i\mathcal{A}_i}(\mathbf{u})=\sum_i h_{\mathcal{A}_i}(\mathbf{u}) for a finite Minkowski sum. Let U={uℓ}ℓ=1K⊂S5\mathcal{U}=\{\mathbf{u}_\ell\}_{\ell=1}^{K}\subset\mathbb{S}^{5} denote KK shared support directions sampling the global wrench space in the object frame. For contact ii and direction uℓ\mathbf{u}_\ell, define

vℓi=G~iTuℓ=[vℓinvℓit],(10)\mathbf{v}_\ell^i=\widetilde{\mathbf{G}}_i^{\mathsf{T}}\mathbf{u}_\ell=\begin{bmatrix}v_{\ell i}^{n}\\\mathbf{v}_{\ell i}^{t}\end{bmatrix}, \tag*{(10)}

where vℓin∈Rv_{\ell i}^{n}\in\mathbb{R} and vℓit∈R2\mathbf{v}_{\ell i}^{t}\in\mathbb{R}^{2} denote the normal and tangential components, respectively.

Maximizing over the local force set gives the single-contact support:

hW~i(uℓ)=hFi(G~iTuℓ)=fn,i(vℓin+μi∥vℓit∥2).(11)\begin{aligned} h_{\widetilde{\mathcal{W}}_i}(\mathbf{u}_\ell)=h_{\mathcal{F}_i}\left(\widetilde{\mathbf{G}}_i^{\mathsf{T}}\mathbf{u}_\ell\right) \\ &=f_{n,i}\left(v_{\ell i}^{n}+\mu_i\left\|\mathbf{v}_{\ell i}^{t}\right\|_2\right). \tag*{(11)} \end{aligned}

Let bℓi=vℓin+μi∥vℓit∥2b_{\ell i}=v_{\ell i}^{n}+\mu_i\left\|\mathbf{v}_{\ell i}^{t}\right\|_2. Because the contact force sets are independent, the support of their wrench sum is additive:

hW~C(uℓ)=∑i=1Mbℓifn,i.(12)h_{\widetilde{\mathcal{W}}_C}(\mathbf{u}_\ell)=\sum_{i=1}^{M}b_{\ell i}f_{n,i}. \tag*{(12)}

Stacking the support values along all KK directions yields

h=[hW~C(u1)⋮hW~C(uK)]=Bfn,[B]ℓi=bℓi.(13)\mathbf{h}=\begin{bmatrix}h_{\widetilde{\mathcal{W}}_C}(\mathbf{u}_1)\\\vdots\\h_{\widetilde{\mathcal{W}}_C}(\mathbf{u}_K)\end{bmatrix}=\mathbf{B}\mathbf{f}_n,\qquad[\mathbf{B}]_{\ell i}=b_{\ell i}. \tag*{(13)}

where fn=[fn,1,…,fn,M]T\mathbf{f}_n = [f_{n,1}, \ldots, f_{n,M}]^\mathsf{T}. Restoring the frame and embodiment indices, the human and robot support profiles are therefore

htH=BtHfn,tH,htR=BtRfn,tR.(14)\mathbf{h}^{\mathrm{H}}_t = \mathbf{B}^{\mathrm{H}}_t\mathbf{f}^{\mathrm{H}}_{n,t}, \qquad\mathbf{h}^{\mathrm{R}}_t = \mathbf{B}^{\mathrm{R}}_t\mathbf{f}^{\mathrm{R}}_{n,t}. \tag*{(14)}

For the human hand, BtH∈RK×MH\mathbf{B}^{\mathrm{H}}_t \in\mathbb{R}^{K \times M_{\mathrm{H}}} is constructed from the contact-aligned MANO geometry ξtc\boldsymbol{\xi}^{\mathrm{c}}_t, while fn,tH\mathbf{f}^{\mathrm{H}}_{n,t} is obtained from the calibrated tactile measurements, with zero force assigned to inactive contacts. For the robot hand, BtR∈RK×MR\mathbf{B}^{\mathrm{R}}_t \in\mathbb{R}^{K \times M_{\mathrm{R}}} is constructed from the contact-aligned configuration qtc\mathbf{q}^{\mathrm{c}}_t, while fn,tR\mathbf{f}^{\mathrm{R}}_{n,t} contains the normal force targets to be optimized. In our implementation, MH=10M_{\mathrm{H}} = 10 and MR=5M_{\mathrm{R}} = 5.

Rather than enforcing exact support matching, we encourage the robot CWS to cover the human CWS over the global direction set U\mathcal{U}. For a candidate robot normal force vector f\mathbf{f}, let σt=max⁡(∥htH∥2/K,ϵ)\sigma_t = \max(\lVert\mathbf{h}^{\mathrm{H}}_t\rVert_2/\sqrt{K}, \epsilon), where ϵ>0\epsilon> 0 is a numerical safeguard. We define the one-sided coverage loss as

Lcoverage(t,f)=12Kσt2∥[htH−BtRf]+∥22,(15)\mathcal{L}_{\mathrm{coverage}}(t,\mathbf{f}) = \frac{1}{2K\sigma_t^2}\left\lVert[\mathbf{h}^{\mathrm{H}}_t-\mathbf{B}^{\mathrm{R}}_t\mathbf{f}]_+\right\rVert_2^2, \tag*{(15)}

where [x]+=max⁡(x,0)[\mathbf{x}]_+ = \max(\mathbf{x},0) is applied elementwise. The loss penalizes only directions in which the robot support is insufficient.

The complete force objective is

Ltotal(t,f)=λcLcoverage(t,f)+λrLratio(t,f)+λtLtemporal(t,f)+λfLforce(t,f).(16)\mathcal{L}_{\mathrm{total}}(t,\mathbf{f}) = \lambda_c\mathcal{L}_{\mathrm{coverage}}(t,\mathbf{f}) + \lambda_r\mathcal{L}_{\mathrm{ratio}}(t,\mathbf{f}) + \lambda_t\mathcal{L}_{\mathrm{temporal}}(t,\mathbf{f}) + \lambda_f\mathcal{L}_{\mathrm{force}}(t,\mathbf{f}). \tag*{(16)}

Here, Lratio\mathcal{L}_{\mathrm{ratio}} preserves the relative force pattern measured at the human fingertips, Ltemporal\mathcal{L}_{\mathrm{temporal}} promotes frame-to-frame continuity, and Lforce\mathcal{L}_{\mathrm{force}} favors smaller forces among solutions with comparable coverage. The nonnegative λc\lambda_c, λr\lambda_r, λt\lambda_t, and λf\lambda_f weight the four terms.

We minimize this convex objective frame by frame using a Nesterov-accelerated projected gradient method. Each iteration projects the solution onto the feasible set defined by constraints on nonnegativity, individual finger forces, total force, and variation between frames. Subsequent frames are initialized from the preceding solution; at t=1t = 1, the temporal term and variation constraint are omitted. Frames with negligible human support or admissible total force are assigned zero target force. The resulting fn,tR∗∈RMR\mathbf{f}^{\mathrm{R}*}_{n,t} \in\mathbb{R}^{M_{\mathrm{R}}} contains the five robot fingertip normal force targets.

With each robot contact normal expressed in the world frame, the desired fingertip force is

Ft,i∗=fn,t,iR∗nt,iW.(17)\mathbf{F}^{*}_{t,i} = f^{\mathrm{R}*}_{n,t,i}\mathbf{n}^{\mathrm{W}}_{t,i}. \tag*{(17)}

The tangential friction component is used to characterize the admissible CWS but is not explicitly recovered during this conversion. The desired fingertip forces induce the generalized finger torque

τt∗=∑i=1MR(Jt,iR)TFt,i∗,(18)\boldsymbol{\tau}^{*}_t = \sum_{i=1}^{M_{\mathrm{R}}}(\mathbf{J}^{\mathrm{R}}_{t,i})^\mathsf{T}\mathbf{F}^{*}_{t,i}, \tag*{(18)}

where Jt,iR\mathbf{J}^{\mathrm{R}}_{t,i} is the translational fingertip Jacobian evaluated at the contact-aligned robot configuration qtc\mathbf{q}^{\mathrm{c}}_t.

Under a quasi-static proportional position control model, τt∗≃KpΔqtw\boldsymbol{\tau}^{*}_t \simeq\mathbf{K}_p\Delta\mathbf{q}^{\mathrm{w}}_t, yielding

Δqtw=Kp−1τt∗,qtw=qtc+Δqtw.(19)\Delta\mathbf{q}^{\mathrm{w}}_t = \mathbf{K}_p^{-1}\boldsymbol{\tau}^{*}_t, \qquad\mathbf{q}^{\mathrm{w}}_t = \mathbf{q}^{\mathrm{c}}_t + \Delta\mathbf{q}^{\mathrm{w}}_t. \tag*{(19)}

All geometry-dependent quantities are evaluated at qtc\mathbf{q}^{\mathrm{c}}_t, such that Δqtw\Delta\mathbf{q}^{\mathrm{w}}_t acts as a local wrench-guided correction. The resulting joint position trajectory {qtw}t=1T\{\mathbf{q}^{\mathrm{w}}_t\}_{t=1}^{T} is finally executed by the robot, while the optimized contact forces serve only as intermediate physical targets.

Experiments

The experiments are organized into three parts. We first describe the acquisition of human demonstrations with synchronized kinematic and tactile measurements, which provide the inputs to the proposed retargeting framework. We then conduct real robot experiments as the primary evaluation of trajectory replayability, task success, and human-to-robot wrench transfer. Finally, simulation experiments provide a complementary evaluation of trajectory feasibility and characterize the gap between simulated and real-world execution.

Human Demonstration Acquisition

Human demonstrations are collected in the optical motion capture workspace shown in Figure 3(a). Hand kinematics are tracked using 21 reflective markers mounted on a tactile glove equipped with 19 sensing channels, while object motion is captured using five reflective markers attached to each manipulated object. The hand motion, object motion, and tactile measurements are synchronized using timestamps to form the aligned kinematic–tactile demonstrations used as input to the proposed framework.

Experimental setups: optical motion capture workspace and UR7e-mounted Sharpa Wave dexterous hand

Figure 3. Experimental setups. (a) Optical motion capture workspace used to collect human demonstrations. (b) UR7e-mounted Sharpa Wave dexterous hand, three payload conditions, and the fixed receiving container used in the real robot pouring experiments.

A single operator performs the pouring task under three payload conditions using the same 22 oz paper cup as the manipulated object: 500 g of rice, 750 g of ceramic beads, and 1500 g of steel beads. We collect 20 demonstrations for each condition, yielding 60 trajectories in total. The demonstrations are acquired at 120 Hz, last approximately 3–8 s, and comprise about 40,000 synchronized frames in total.

Real Robot Experiments

Real-world evaluations are conducted using a Sharpa Wave dexterous hand mounted on a UR7e robot arm, as shown in Figure 3(b). During replay, the synchronized arm–hand trajectory targets are updated at 120 Hz. Both the UR7e and Sharpa Wave controllers operate at 100 Hz and apply the latest available targets at each control cycle, while tactile measurements are acquired at 30 Hz for each finger. For hardware safety, all trajectories are replayed at 0.2×0.2\times the original demonstration speed. We execute the generated joint position trajectories on the physical system without online policy correction to test whether the retargeted interactions can be realized on hardware.

PayloadMethodReplay success ↑\uparrowGlobal wrench space: Covering rate CC (%) ↑\uparrowGlobal wrench space: Appropriate rate AA (%) ↑\uparrowGlobal wrench space: Margin MM (→1\to1)Task-relevant wrench space: Covering rate CC (%) ↑\uparrowTask-relevant wrench space: Appropriate rate AA (%) ↑\uparrowTask-relevant wrench space: Margin MM (→1\to1)
RiceDexPilot with preload15/20 (75%)32.3 ±\pm 20.629.7 ±\pm 18.00.8 ±\pm 0.445.7 ±\pm 30.641.7 ±\pm 28.91.1 ±\pm 0.7
RiceManipTrans0/20 (0%)N/AN/AN/AN/AN/AN/A
RiceOurs19/20 (95%)52.9 ±\pm 15.940.0 ±\pm 10.61.6 ±\pm 1.560.0 ±\pm 23.637.6 ±\pm 28.51.9 ±\pm 0.9
Ceramic beadsDexPilot with preload17/20 (85%)49.8 ±\pm 23.934.7 ±\pm 14.61.7 ±\pm 1.760.5 ±\pm 29.934.3 ±\pm 24.32.5 ±\pm 3.0
Ceramic beadsManipTrans0/20 (0%)N/AN/AN/AN/AN/AN/A
Ceramic beadsOurs19/20 (95%)47.4 ±\pm 19.239.7 ±\pm 13.01.3 ±\pm 0.757.7 ±\pm 29.241.1 ±\pm 26.71.8 ±\pm 1.6
Steel beadsDexPilot with preload16/20 (80%)56.2 ±\pm 21.536.5 ±\pm 14.62.3 ±\pm 2.869.6 ±\pm 30.237.8 ±\pm 26.73.6 ±\pm 5.0
Steel beadsManipTrans0/20 (0%)N/AN/AN/AN/AN/AN/A
Steel beadsOurs20/20 (100%)53.2 ±\pm 14.642.7 ±\pm 11.61.4 ±\pm 0.767.7 ±\pm 26.348.8 ±\pm 25.02.0 ±\pm 1.3

Table I. Quantitative real robot evaluation across three pouring payloads. Success is reported as successful trials out of 20 (percentage), and wrench metrics are reported as mean ±\pm standard deviation. The best value for each payload and metric is shown in bold.

We consider three pouring conditions by loading the same 22 oz source cup with 500 g of rice, 750 g of ceramic beads, or 1500 g of steel beads. The receiving container is rigidly fixed to the workbench using metal angle brackets to maintain a consistent target location across trials. The dexterous hand is mounted with a 51 mm translational offset along the zz-axis of the UR7e tool center point (TCP) frame. The arm motion is taught separately and synchronized with the retargeted hand trajectory, after which the coupled arm–hand motion executes four phases: grasp and lift, tilt, pour, and hold. To establish a repeatable spatial registration between the robot and the experimental scene, we mount a custom alignment fixture that matches the footprint of the source cup to the UR7e end effector. Using kinesthetic teaching, we register the cup and receiving container locations on the workbench, while metal angle brackets physically constrain their placement across trials. The resulting scene geometry is expressed in the robot base frame through rigid coordinate transformations.

For each replay, the dexterous hand is first moved to the registered cup pose and commanded to execute the first-frame grasp configuration. After establishing the grasp, the arm lifts the cup and moves it to the initial pose of the pouring trajectory, which is transformed with respect to the registered receiving container frame. The synchronized arm–hand trajectory is then replayed from this aligned initial state. We conduct 20 real robot trials for each method–payload pair. A trial is considered successful if the cup remains securely grasped throughout the complete trajectory, exhibits no substantial deformation or loss of grasp stability, and the payload remains contained without unintended spillage. Task performance is reported as the number of successful executions among the 20 trials. Figure 4 shows representative frames from a human pouring demonstration and the corresponding robot

Qualitative visualization of a representative pouring sequence

Figure 4. Qualitative visualization of a representative pouring sequence. (a) Human demonstration and (b) the corresponding robot replay across four task phases: grasp and lift, tilt, pour, and hold.

replay across the four task phases: grasp and lift, tilt, pour, and hold.

To evaluate interaction fidelity, we compute metrics based on support functions over both the global and task-relevant wrench spaces. The global evaluation queries directions over the normalized six-dimensional wrench space. The task-relevant wrench space evaluation separately considers a force cone centered on the object axis most closely aligned with gravity and a torque cone centered on the axis connecting the fingertips of the human thumb and middle finger; the two sets of metrics are then averaged with equal weight to obtain the task-relevant wrench space macro results. For the real-robot evaluation, the robot support is computed from the fingertip normal forces measured during replay. At each frame and within each queried direction set, let HkH_k and RkR_k denote the human and robot support values along direction kk, respectively, and define V={k∣Hk>ε}V = \{k \mid H_k > \varepsilon\} and ρk=Rk/Hk\rho_k = R_k/H_k for k∈Vk \in V, where ε=10−6\varepsilon= 10^{-6}. We report three metrics: (i) covering rate, C=∣{k∈K∣Rk≥Hk−ε}∣/KC = |\{k \in K \mid R_k \ge H_k - \varepsilon\}|/K, is the fraction of all KK directions in which the robot support meets or exceeds the human reference; (ii) appropriate rate, A=∣{k∈V∣1≥ρk≥τ}∣/∣V∣A = |\{k \in V \mid1 \ge\rho_k \ge\tau\}|/|V|, is the fraction of valid directions providing sufficient but not excessive support, with U=3U = 3; and (iii) margin, M=∣V∣−1∑k∈VρkM = |V|^{-1}\sum_{k\in V}\rho_k, is the dimensionless mean support ratio, where values below, equal to, or above one indicate insufficient, matched, or redundant support on average. The reported mean and standard deviation are computed over all available per-frame metric values, with equal weight assigned to each frame.

We compare the complete framework against two baselines. DexPilot with preload augments DexPilot kinematic retargeting [6] with a fixed contact preload to improve direct physical replay. Starting from the kinematically retargeted hand pose, we project each fingertip tactile patch center onto the cup mesh and define the preload direction from the fingertip toward the corresponding projected point. A 15 N preload is then applied along this direction. The resulting fingertip forces are mapped to generalized joint torques and subsequently converted into joint position offsets using the joint stiffness KpK_p, yielding the final preload-adjusted hand trajectory.

We reproduce ManipTrans [9] following its original method and training protocol. Its two-stage pipeline first pretrains a generalist trajectory imitator to track human hand motions and then fine-tunes a task-specific residual module under interaction constraints. For hardware evaluation, we export the joint trajectories from the simulated rollouts and replay them open loop using the same physical setup as the other methods. Table I summarizes the quantitative comparison.

Across the three payload conditions, our method succeeds in 58 of 60 trials (96.7%), whereas DexPilot with preload succeeds in 48 of 60 trials (80.0%), representing an improvement of 16.7 percentage points. ManipTrans fails all 60 real robot trials. Our method achieves the best value in 11 of the 18 continuous wrench metrics. It improves the global appropriate rate under all three payload conditions and obtains higher task-relevant wrench space appropriate rates and margins closer to unity for both ceramic and steel beads. DexPilot with preload remains stronger on several covering metrics and on several rice-specific measures. The preload-augmented DexPilot baseline serves as a combined ablation of the two physics-aware stages, assessing their joint benefit over kinematic retargeting under a fixed 15 N contact preload.

Because none of the ManipTrans trajectories yields a valid real robot execution, its wrench transfer metrics are reported as N/A. In particular, wrench fidelity cannot be meaningfully evaluated when the manipulated cup is dropped or substantially deformed, or when the intended pouring sequence is not completed. The difference between its simulation and real-world performance suggests a substantial sim-to-real discrepancy under our deployment conditions, particularly with cup deformation and heavier payloads.

Simulation Experiments

Simulation provides a complementary test of trajectory feasibility. We evaluate our method and DexPilot with preload in MuJoCo and roll out the complete ManipTrans trajectory in its original Isaac Gym setting. This protocol retains each method’s native simulator. For each material, 20 distinct trajectories are each rolled out once. A rollout is successful if the manipulated cup is not dropped throughout the complete trajectory; accordingly, we report only success rates in Table II.

MaterialDexPilot with preloadManipTransOurs
Rice20/20 (100%)20/20 (100%)20/20 (100%)
Ceramic beads20/20 (100%)12/20 (60%)20/20 (100%)
Steel beads19/20 (95%)11/20 (55%)20/20 (100%)

Table II. Simulation success rates under three payload conditions.

Our method succeeds on all 60 simulated trajectories, while DexPilot with preload succeeds on 59 of 60 (98.3%). ManipTrans succeeds on 43 of 60 trajectories (71.7%): its success rate is 100% for rice but decreases to 60% and 55% for ceramic and steel beads, respectively. The drop from 71.7% success in simulation to 0% on hardware indicates a substantial sim-to-real gap under our deployment conditions. Because our method and the preload baseline both approach the simulation ceiling, the real robot experiments provide the more discriminative evaluation.

Conclusion

This letter presented a hierarchical three-stage retargeting framework that jointly aligns kinematic and tactile information across human and robot hands. The framework uses object-centric contact wrench space as a shared physical representation for tactile alignment, complementing the alignment of hand kinematics and contact geometry. By converting optimized fingertip force targets into joint position corrections, it produces robot hand trajectories for direct offline replay without online policy correction. In paper cup pouring experiments under three payload conditions, our method achieved a 96.7% success rate over 60 real robot trials and improved wrench transfer fidelity in both global and task-relevant wrench spaces. The resulting robot demonstrations retain human-derived wrench support, providing a practical data source for future contact-rich manipulation learning. The current evaluation is limited to a single operator, a fixed workspace, and one family of pouring tasks. Future work will extend the framework to in-hand manipulation and tool use.

References

  1. [1]A. Rajeswaran, V. Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, and S. Levine, “Learning Complex Dexterous Manipulation with Deep Reinforcement Learning and Demonstrations,” in Proceedings of Robotics: Science and Systems. Pittsburgh, Pennsylvania: Robotics: Science and Systems Foundation, 2018. [Online]. Available: https://www.roboticsproceedings.org/rss14/p49.htmlarxiv.org/abs/1709.10087
  2. [2]C. Wang, H. Shi, W. Wang, R. Zhang, L. Fei-Fei, and C. K. Liu, “DexCap: Scalable and Portable Mocap Data Collection System for Dexterous Manipulation,” in Proceedings of Robotics: Science and Systems. Robotics: Science and Systems Foundation, 2024. [Online]. Available: https://www.roboticsproceedings.org/rss20/p043.htmlarxiv.org/abs/2403.07788
  3. [3]R. Meattini, R. Suárez, G. Palli, and C. Melchiorri, “Human to Robot Hand Motion Mapping Methods: Review and Classification,” IEEE Transactions on Robotics, vol. 39, no. 2, pp. 842–861, 2023. [Online]. Available: https://ieeexplore.ieee.org/document/9896989/DOI
  4. [4]Y. Qin, H.-W. Wu, S. Liu, H. Jiang, R. Yang, Y. Fu, and X. Wang, “DexMV: Imitation Learning for Dexterous Manipulation from Human Videos,” in Computer Vision – ECCV 2022, ser. Lecture Notes in Computer Science, vol. 13699. Cham: Springer Nature Switzerland, 2022, pp. 570–587. [Online]. Available: https://link.springer.com/chapter/10.1007/978-3-031-19842-7_33arxiv.org/abs/2108.05877
  5. [5]C. Chen, Z. Yu, H. Choi, M. Cutkosky, and J. Bohg, “DexForce: Extracting Force-Informed Actions from Kinesthetic Demonstrations for Dexterous Manipulation,” IEEE Robotics and Automation Letters, vol. 10, no. 6, pp. 6416–6423, 2025. [Online]. Available: https://ieeexplore.ieee.org/document/10993295/arxiv.org/abs/2501.10356
  6. [6]A. Handa, K. Van Wyk, W. Yang, J. Liang, Y.-W. Chao, Q. Wan, S. Birchfield, N. Ratliff, and D. Fox, “DexPilot: Vision-Based Teleoperation of Dexterous Robotic Hand-Arm System,” in 2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 9164–9170. [Online]. Available: https://ieeexplore.ieee.org/document/9197124/arxiv.org/abs/1910.03135
  7. [7]A. Lakshmipathy, D. Bauer, C. Bauer, and N. S. Pollard, “Contact Transfer: A Direct, User-Driven Method for Human to Robot Transfer of Grasps and Manipulations,” in 2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022, pp. 6195–6201. [Online]. Available: https://ieeexplore.ieee.org/document/9811739/arxiv.org/abs/2110.15532
  8. [8]J. Wu, S. Yao, G. He, X. Liu, Z. Zeng, X. Jiang, H. Yang, W. Zhang, and H. Zhao, “TopoRetarget: Interaction-Preserving Retargeting for Dexterous Manipulation,” 2026, arXiv:2606.16272; accepted to the Conference on Robot Learning (CoRL 2026). [Online]. Available: https://arxiv.org/abs/2606.16272
  9. [9]K. Li, P. Li, T. Liu, Y. Li, and S. Huang, “ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2025, pp. 6991–7003. [Online]. Available: https://ieeexplore.ieee.org/document/11094147/arxiv.org/abs/2503.21860
  10. [10]S. Zhao, X. Zhu, Y. Chen, C. Li, Y. Xie, X. Zhang, M. Ding, and M. Tomizuka, “DexH2R: Task-Oriented Dexterous Manipulation from Human to Robots,” IEEE/ASME Transactions on Mechatronics, vol. 31, no. 3, pp. 3202–3213, 2026. [Online]. Available: https://ieeexplore.ieee.org/document/11316901/arxiv.org/abs/2411.04428
  11. [11]M. Zhao, Y. Hou, D. Fox, Y. Narang, A. Mandlekar, and S. Song, “DexMachina: Functional Retargeting for Bimanual Dexterous Manipulation,” in Proceedings of the 43rd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 306. PMLR, 2026, pp. 86198–86213. [Online]. Available: https://proceedings.mlr.press/v306/mandi26a.htmlarxiv.org/abs/2505.24853
  12. [12]Y.-W. Chao, W. Yang, Y. Xiang, P. Molchanov, A. Handa, J. Tremblay, Y. S. Narang, K. Van Wyk, U. Iqbal, S. Birchfield, J. Kautz, and D. Fox, “DexYCB: A Benchmark for Capturing Hand Grasping of Objects,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2021, pp. 9040–9049. [Online]. Available: https://ieeexplore.ieee.org/document/957876/arxiv.org/abs/2104.04631
  13. [13]Z. Fan, O. Taheri, D. Tzionas, M. Kocabas, M. Kaufmann, M. J. Black, and O. Hilliges, “ARCTIC: A Dataset for Dexterous Bimanual Hand-Object Manipulation,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023, pp. 12 943–12 954. [Online]. Available: https://ieeexplore.ieee.org/document/10203858/arxiv.org/abs/2204.13662
  14. [14]H. Pan, K. Zhu, D. Li, Y. Chen, S. Jiang, B. He, and P. B. Shull, “Decoding Human Touch Noninvasively: Tactile Inference From EMG and Kinematics Using AET-TacNet,” IEEE Transactions on Industrial Informatics, vol. 22, no. 6, pp. 5117–5128, 2026. [Online]. Available: https://doi.org/10.1109/TII.2026.3665344
  15. [15]S. Ni, H. Zhang, Z. Wei, G. Chen, C. Zhang, Y. Shi, and J. Wang, “TactiDex: A Real-World Tactile-Guided Benchmark for Human-Like Dexterous Manipulation,” 2026, arXiv:2607.09190; accepted to the 34th ACM International Conference on Multimedia (ACM MM 2026). [Online]. Available: https://arxiv.org/abs/2607.09190
  16. [16]Y. Qin, H. Su, and X. Wang, “From One Hand to Multiple Hands: Imitation Learning for Dexterous Manipulation From Single-Camera Teleoperation,” IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 10 873–10 881, 2022. [Online]. Available: https://ieeexplore.ieee.org/document/9849105/arxiv.org/abs/2204.12490
  17. [17]Z.-H. Yin, C. Wang, L. Pineda, K. Bodduluri, T. Wu, P. Abbeel, and M. Mukadam, “Geometric Retargeting: A Principled, Ultrafast Neural Hand Retargeting Algorithm,” in 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Hangzhou, China: IEEE, 2025, pp. 17 376–17 382. [Online]. Available: https://ieeexplore.ieee.org/document/11247700/
  18. [18]X. Zhu, Z. Liu, S. Jain, C. Li, M. Noori, M. A. Lin, H. Zhao, J. Welsh, M. Verghese, W. Liu, T. Wang, X. Da, Z. Luo, V. Kulkarni, N. Bhatti, Y. Zhu, L. Fan, B. Wen, D. Xu, S. Pouya, and Y. Chang, “Learning Dexterous Manipulation Using Contact Wrench Guidance From Human Demonstration,” 2026, arXiv:2607.00033; accepted to the Conference on Robot Learning (CoRL) 2026. [Online]. Available: https://arxiv.org/abs/2607.00033
  19. [19]C. Pan, C. Wang, H. Qi, Z. Liu, H. Bharadhwaj, A. Sharma, T. Wu, G. Shi, J. Malik, and F. Hogan, “SPIDER: Scalable Physics-Informed Dexterous Retargeting,” 2026, arXiv:2511.09484; accepted to the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2026). [Online]. Available: https://arxiv.org/abs/2511.09484
  20. [20]C. Ferrari and J. Canny, “Planning optimal grasps,” in Proceedings 1992 IEEE International Conference on Robotics and Automation. Nice, France: IEEE Comput. Soc. Press, 1992, pp. 2290–2295. [Online]. Available: http://ieeexplore.ieee.org/document/219918/
  21. [21]X. Zhu, H. Ding, and J. Wang, “Grasp analysis and synthesis based on a new quantitative measure,” Robotics and Automation, IEEE Transactions on, vol. 19, pp. 942–953, 2004.DOI
  22. [22]M. A. Roa and R. Suárez, “Grasp quality measures: Review and performance,” Autonomous Robots, vol. 38, no. 1, pp. 65–88, 2015. [Online]. Available: http://link.springer.com/10.1007/s10514-014-9402-3pubmed.ncbi.nlm.nih.gov/26074671
  23. [23]T. Liu, Z. Liu, Z. Jiao, Y. Zhu, and S.-C. Zhu, “Synthesizing Diverse and Physically Stable Grasps with Arbitrary Hand Structures using Differentiable Force Closure Estimator,” IEEE Robotics and Automation Letters, vol. 7, no. 1, pp. 470–477, 2022. [Online]. Available: https://ieeexplore.ieee.org/document/9619920/DOI
  24. [24]R. Wang, J. Zhang, J. Chen, Y. Xu, P. Li, T. Liu, and H. Wang, “DexGraspNet: A Large-Scale Robotic Dexterous Grasp Dataset for General Objects Based on Simulation,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 11 359–11 366. [Online]. Available: https://ieeexplore.ieee.org/document/10160982/arxiv.org/abs/2210.02697
  25. [25]J. Chen, Y. Chen, J. Zhang, and H. Wang, “Task-Oriented Dexterous Hand Pose Synthesis Using Differentiable Grasp Wrench Boundary Estimator,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 5281–5288. [Online]. Available: https://ieeexplore.ieee.org/document/10802652/arxiv.org/abs/2309.13586
  26. [26]J. Chen, Y. Ke, and H. Wang, “BODex: Scalable and Efficient Robotic Dexterous Grasp Synthesis Using Bilevel Optimization,” in 2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 1–8. [Online]. Available: https://ieeexplore.ieee.org/document/11127930/arxiv.org/abs/2412.16490
  27. [27]J. Chen, Y. Ke, L. Peng, and H. Wang, “Dexonomy: Synthesizing All Dexterous Grasp Types in a Grasp Taxonomy,” in Proceedings of Robotics: Science and Systems. Robotics: Science and Systems Foundation, 2025. [Online]. Available: https://www.roboticsproceedings.org/rss21/p105.htmlarxiv.org/abs/2504.18829
  28. [28]J. Romero, D. Tzionas, and M. J. Black, “Embodied hands: Modeling and capturing hands and bodies together,” ACM Transactions on Graphics, vol. 36, no. 6, pp. 245:1–245:17, 2017. [Online]. Available: https://dl.acm.org/doi/10.1145/3130800.3130883DOI
  29. [29]G. Pavlakos, D. Shan, I. Radosavovic, A. Kanazawa, D. Fouhey, and J. Malik, “Reconstructing hands in 3D with transformers,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2024, pp. 9826–9836. [Online]. Available: https://ieeexplore.ieee.org/document/10655481/

Paper details

vlnc:W_dw5r8my8
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