Interaction-Preserving Motion Retargeting

Updated 17 July 2026

Interaction-preserving motion retargeting is a method that transfers motion between source and target embodiments by preserving semantic interaction cues rather than replicating exact joint trajectories.
It employs geometry-, contact-, and physics-aware representations to maintain critical structures such as hand-object and foot-ground contacts across diverse morphologies.
Approaches range from kinematic optimization and learning-based techniques to physics-based simulation, balancing interaction fidelity with dynamic feasibility.

Interaction-preserving motion retargeting is the transfer of motion from a source embodiment to a target embodiment under the constraint that the target should preserve interaction-relevant structure rather than merely reproduce joint trajectories. Across the literature, the preserved structure may be self-contact, hand-object contact, foot-ground contact, support phases, object-relative geometry, multi-character spatial relations, or, in a narrower dynamics-oriented sense, internal coordination patterns induced by morphology and inertia. The central distinction is between literal pose equivalence and preservation of the relations that make the motion semantically or physically meaningful under embodiment change. Several recent works therefore formulate retargeting around contact timing, interaction meshes, proximity fields, end-effector semantics, or physics-based feasibility rather than around skeleton-space imitation alone (Reda et al., 2023, Wu et al., 15 Jun 2026, Yang et al., 30 Sep 2025).

1. Definition, scope, and problem formulations

Interaction-preserving motion retargeting appears in several distinct regimes. In dexterous manipulation, the source is a human hand-object demonstration and the target is a robot hand trajectory that must preserve task-relevant hand-object interaction rather than only fingertip pose (Wu et al., 15 Jun 2026, Lakshmipathy et al., 2024). In humanoid loco-manipulation, the source may be a human demonstration involving terrain, objects, and scene contact, and the target is a humanoid robot trajectory that must preserve body-object-environment relationships while satisfying kinematic constraints (Yang et al., 30 Sep 2025). In sparse-input avatar control, the source signal may be only headset and controller poses, and the target is a physically simulated character whose motion should preserve user intent, gait timing, balance-relevant behavior, and foot-ground contact semantics despite severe morphology mismatch (Reda et al., 2023). In multi-character settings, the source is an interaction clip and the target is a physically simulated multi-agent system that should preserve the relative geometry defining the interaction (Zhang et al., 2023).

The scope of “interaction” varies accordingly. Some papers define it narrowly as support/contact structure, especially foot-ground contact and stance timing in locomotion-oriented retargeting (Huang et al., 2 Jun 2026, Reda et al., 2023). Others broaden it to self-contact, near-body proximity, self-penetration avoidance, and inter-character contact in skinned character animation (Cheynel et al., 28 Feb 2025, Choi et al., 19 May 2026, Ye et al., 2024, Villegas et al., 2021). Manipulation-focused work treats interaction as the evolving local structure of hand-object contact regions, object-relative geometry, and contact transitions over time (Wu et al., 15 Jun 2026, Lakshmipathy et al., 2024, Feng et al., 13 Jul 2026). A different line interprets interaction as internal dynamical coupling among joints and preserves piecewise geodesic synergy structure under a Riemannian metric induced by the mass-inertia matrix rather than explicit external contact (Klein et al., 2022).

A recurring claim across these formulations is that preserving interaction semantics often requires not preserving the literal source pose. Physics-based sparse-input retargeting explicitly notes that when morphology mismatch is large, the system preserves root progression, contact timing, and orientation semantics rather than exact lower-body geometry (Reda et al., 2023). Human-to-humanoid retargeting with unpaired data similarly preserves body-environment interaction structure mainly through end-effector semantics and foot-contact timing, not exact pose matching (Huang et al., 2 Jun 2026). For dexterous manipulation, contact-rich methods argue that visually plausible hand posture is insufficient if local hand-object interaction topology is lost, because downstream RL reference tracking then degrades (Wu et al., 15 Jun 2026, Feng et al., 13 Jul 2026).

2. Representations of interaction

The field is distinguished by the choice of interaction representation. Skeleton-only retargeting typically fails when interaction depends on surface geometry, object-relative placement, or feasibility under embodiment change. Interaction-preserving approaches therefore introduce geometric, contact-aware, or physics-aware representations that are more invariant to morphology while remaining task-relevant.

Representation	Interaction captured	Representative papers
Contact areas / contact pairs	Hand-object surface contacts over full manipulations	(Lakshmipathy et al., 2024)
Interaction graph / interaction mesh	Local geometry among body, object, terrain, or multiple agents	(Wu et al., 15 Jun 2026, Feng et al., 13 Jul 2026, Yang et al., 30 Sep 2025, Zhang et al., 2023)
Proximity fields / adaptive anchors / dense mesh interaction	Self-contact, near-body proximity, contact preservation, interpenetration avoidance	(Cheynel et al., 28 Feb 2025, Choi et al., 19 May 2026, Ye et al., 2024, Villegas et al., 2021)
End-effector and contact timing semantics	Foot-ground contacts, stance timing, morphology-invariant end-effector behavior	(Huang et al., 2 Jun 2026)
Dynamics-induced manifold structure	Internal joint coordination as geodesic synergies	(Klein et al., 2022)

For dexterous manipulation, one influential representation is the sparse interaction graph over hand keypoints and object surface samples. TopoRetarget defines source and robot vertex sets as

$V_t^s = [P_t^h; O_t], \qquad V_t^r(q) = [P_t^r(q); O_t],$

constructs shared connectivity through Delaunay tetrahedralization, and preserves weighted Laplacian coordinates of the resulting graph so that local hand-object and intra-hand neighborhoods are preserved under embodiment change (Wu et al., 15 Jun 2026). REGRIND adopts a closely related interaction-mesh formulation for hand-object motion, with Delaunay tetrahedralization over human hand keypoints and object keypoints, and minimizes deformation of Laplacian coordinates subject to joint limits, velocity bounds, and non-penetration constraints (Feng et al., 13 Jul 2026). OmniRetarget extends the interaction-mesh idea to whole-body humanoid retargeting by including body keypoints together with sampled object and environment points, and preserving the coupled body-object-environment structure through Laplacian deformation (Yang et al., 30 Sep 2025).

In skinned character retargeting, the representation often shifts from sparse skeletons to sparse or dense surface anchors. ReConForM uses rigged key vertices transferred from a generic humanoid template through entropy-regularized optimal transport and extracts pairwise distance, direction, penetration, height-to-ground, and sliding descriptors from their trajectories (Cheynel et al., 28 Feb 2025). The weighting coefficients

$\mathcal{W}_\text{interaction}(i,j)=\text{clamp}\left(1-\frac{\mathcal{M_\text{dist}(i,j)-d_\text{min}}{d_\text{max}-d_\text{min}}\right), \qquad \mathcal{W}_\text{floor}(i)=\text{clamp}\left(1-\frac{\mathcal{M}_\text{height}(i)-h_\text{min}}{h_\text{max}-h_\text{min}}\right)$

make interactions near contact and near the ground dominant in the objective (Cheynel et al., 28 Feb 2025). MeshRet instead uses semantically consistent sensors and a dense mesh interaction field, with pairwise interaction features

$\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$

so that local directional geometry between body regions is preserved rather than only distances (Ye et al., 2024). Skinned motion retargeting with spatially adaptive interaction guidance replaces static correspondences with spatially adaptive anchors that are repositioned to reachable target regions through Transformer-based refinement and differentiable soft projection, specifically to preserve self-contact and near-body proximity under exaggerated target geometry (Choi et al., 19 May 2026).

Another representational strategy is to encode semantics through normalized end-effector behavior. Human2Humanoid introduces a morphology-invariant end-effector consistency loss that matches normalized end-effector displacements from each embodiment’s T-pose: $\text{normalized displacement} = \frac{\operatorname{FK}_k(q,t)-\operatorname{FK}_k(q^{\mathit{Tpose}})}{S}.$ This captures semantics of reaching and locomotion while removing scale differences, and is complemented by source-derived foot-contact timing (Huang et al., 2 Jun 2026).

3. Methodological families

A first family is purely kinematic optimization with explicit interaction constraints. Kinematic motion retargeting for contact-rich anthropomorphic manipulation formulates retargeting as non-isometric shape matching over dense contact areas, transferring source hand-object contact points through local charts and then solving inverse kinematics with marker, contact, table-penetration, and prior terms: $\theta_f^* = \argmin_{\theta} \quad \lambda_{m} \Gamma_{M} + \lambda_{c} \Gamma_{C} + + \lambda_{t} \Gamma_{T} + \lambda_{j} \Gamma_{J}$ subject to joint limits (Lakshmipathy et al., 2024). Contact-aware retargeting of skinned motion similarly builds a geometry-aware objective over mesh contacts and self-interpenetration,

$E_{\mathit{full}} = E_{\mathit{geo}} + E_{\mathit{skel}},$

with explicit self-contact, interpenetration, and foot-contact terms, then refines motion through geometry-conditioned recurrent inference and encoder-space optimization (Villegas et al., 2021). ReConForM remains kinematic but reduces the problem to sparse mesh descriptors and batch optimization, achieving real-time retargeting while explicitly modeling self-contact, floor contact, inter-character contact, and terrain interaction (Cheynel et al., 28 Feb 2025).

A second family is learning-based skinned-motion retargeting guided by geometric interaction descriptors. R $^2$ ET uses a skeleton-aware residual module to preserve motion semantics and a shape-aware residual module to reduce interpenetration and contact-missing, then blends the two through a balancing gate (Zhang et al., 2023). MeshRet directly models dense geometric interactions in a single stage by aligning the dense mesh interaction field rather than performing skeleton retargeting plus geometry correction (Ye et al., 2024). Spatially adaptive anchor methods push this further by adapting the correspondence structure itself, arguing that fixed source-target anchor mappings fail when the target has exaggerated local morphology (Choi et al., 19 May 2026). This suggests a broader trend away from fixed skeleton-space correspondences toward representations whose support points can move on the target surface when the original semantic region is unreachable.

A third family is physics-based retargeting. Physics-based motion retargeting from sparse inputs trains a torque policy in simulation with reinforcement learning, where the runtime controller receives sparse user observations and simulator state and produces torques, while a rough kinematic retargeting is used only at training time as reward supervision (Reda et al., 2023). DynaRetarget begins from imperfect kinematic humanoid-object trajectories and uses Sampling-Based Trajectory Optimization to refine them into dynamically feasible motions under full simulation, preserving manipulation outcomes through dense robot and object tracking rather than symbolic contact planning (Dhedin et al., 6 Feb 2026). In dexterous manipulation, REGRIND combines interaction-preserving geometric retargeting with residual RL, using the retargeted motion both as a reference trajectory and as a restart-state distribution for learning (Feng et al., 13 Jul 2026). In multi-character interaction, deep RL learns policies that imitate not only each character’s motion but also the interaction itself through an interaction-graph reward (Zhang et al., 2023).

A fourth family addresses whole-body robot teleoperation and contact-rich loco-manipulation through constrained optimization. Multi-contact motion retargeting using whole-body optimization and sequential force equilibrium treats the source as operator end-effector and contact-state commands rather than prerecorded motion. The retargeting layer optimizes whole-body configuration and contact-force distributions under quasi-static equilibrium, contact maintenance, friction, center-of-pressure, and joint/torque constraints, preserving the intended multi-contact interaction as far as feasibility allows (Rouxel et al., 2022). OmniRetarget occupies an adjacent but more data-generation-oriented position: it retargets human motions to humanoids with hard constraints on signed distances, joint limits, velocity limits, and stance-foot sticking while preserving an interaction mesh over the robot, terrain, and manipulated objects (Yang et al., 30 Sep 2025).

A fifth family uses representation learning for cross-morphology motion generation without explicit object interaction. AdaMorph maps human motion into a morphology-agnostic latent intent space, conditions decoding with robot prompts and Adaptive Layer Normalization, and enforces orientation and trajectory consistency through differentiable integration (Zhang et al., 12 Jan 2026). Human2Humanoid learns unpaired human-to-humanoid translation with CycleGAN-style generators, skeleton-aware graph convolution, morphology-invariant end-effector consistency, and explicit foot-contact and foot-height penalties (Huang et al., 2 Jun 2026). These systems preserve dynamics, rhythm, and support-phase structure more than explicit object interaction.

4. Optimization, control, and learning objectives

The optimization targets reveal what each method treats as an invariant. In graph- and mesh-based methods, the central quantity is usually a local geometric descriptor. TopoRetarget preserves the weighted Laplacian coordinates

$\Delta_t(V)_i = \sum_{j\in N_t(i)} w_{ij,t} (v_i - v_j),$

with interaction-mesh energy

$E_{\mathrm{IM}}(q) = \sum_{i=1}^{N_v} \left\| \Delta_t(V_t^r(q))_i - \Delta_t(V_t^s)_i \right\|_2^2,$

and combines this with a bone-direction prior and signed-distance penetration constraints (Wu et al., 15 Jun 2026). REGRIND uses a similar Laplacian deformation term over a combined hand-object mesh, plus temporal smoothness and hard feasibility constraints, to obtain interaction-consistent reference trajectories for RL (Feng et al., 13 Jul 2026). OmniRetarget generalizes the Laplacian formulation to body-object-environment meshes and solves

$q_t^\star = \argmin_{q_t} \; \sum_{i} \|L(p_{t,i} ^{\text{source}})-L(p_{t,i}^{\text{target}}(q_t))\|^2 + \|q_t - q_{t-1}\|_{Q}^2$

subject to non-penetration, joint limits, velocity limits, and stance-foot sticking (Yang et al., 30 Sep 2025).

Contact-aware skinned retargeting often uses direct geometric penalties rather than graph descriptors. Contact-aware retargeting of skinned motion minimizes source-transferred vertex-pair distances for self-contact,

adds a foot-contact term

and combines them with mesh-interpenetration penalties and skeleton-space motion losses (Villegas et al., 2021). ReConForM’s semantic loss explicitly matches contact-sensitive mesh descriptors, including distance, direction, height, and sliding, under proximity-driven weights (Cheynel et al., 28 Feb 2025). Spatially adaptive anchor methods match source and target anchor distance and direction fields while regularizing adapted anchors for reachability and ordering: $\mathcal{W}_\text{interaction}(i,j)=\text{clamp}\left(1-\frac{\mathcal{M_\text{dist}(i,j)-d_\text{min}}{d_\text{max}-d_\text{min}}\right), \qquad \mathcal{W}_\text{floor}(i)=\text{clamp}\left(1-\frac{\mathcal{M}_\text{height}(i)-h_\text{min}}{h_\text{max}-h_\text{min}}\right)$2 and

(Choi et al., 19 May 2026).

Physics-based systems shift the invariant from kinematic closeness to physically realizable behavior. In sparse-input retargeting, the policy controls joint torques in simulation and is trained with

where contact reward matches human foot-contact timing and action reward regularizes smoothness and energy (Reda et al., 2023). DynaRetarget uses forward simulation inside trajectory optimization and heavily weights object pose and velocity tracking, so preservation of interaction is judged by whether the physically rolled-out humanoid-object trajectory remains close to the reference (Dhedin et al., 6 Feb 2026). Multi-contact retargeting for teleoperation enforces quasi-static equilibrium directly,

making contact maintenance and force feasibility hard constraints rather than soft heuristics (Rouxel et al., 2022).

Learning frameworks also differ in whether interaction preservation is explicit or indirect. Human2Humanoid uses explicit source-derived contact gating,

then penalizes target foot velocity and foot height during source-inferred stance (Huang et al., 2 Jun 2026). AdaMorph, by contrast, preserves orientation evolution and trajectory consistency through integration-based losses rather than contact constraints: $\mathcal{W}_\text{interaction}(i,j)=\text{clamp}\left(1-\frac{\mathcal{M_\text{dist}(i,j)-d_\text{min}}{d_\text{max}-d_\text{min}}\right), \qquad \mathcal{W}_\text{floor}(i)=\text{clamp}\left(1-\frac{\mathcal{M}_\text{height}(i)-h_\text{min}}{h_\text{max}-h_\text{min}}\right)$7 (Zhang et al., 12 Jan 2026). This places it closer to semantics- and dynamics-preserving retargeting than to full interaction-aware retargeting.

5. Evaluation regimes and empirical patterns

The empirical evidence in the literature is organized around the interaction regime. For dexterous manipulation, TopoRetarget evaluates on ContactPose using contact precision, contact alignment, maximum penetration depth, fraction of frames with penetration greater than $\mathcal{W}_\text{interaction}(i,j)=\text{clamp}\left(1-\frac{\mathcal{M_\text{dist}(i,j)-d_\text{min}}{d_\text{max}-d_\text{min}}\right), \qquad \mathcal{W}_\text{floor}(i)=\text{clamp}\left(1-\frac{\mathcal{M}_\text{height}(i)-h_\text{min}}{h_\text{max}-h_\text{min}}\right)$8 mm, and solve time. It reports 7.71 mm contact precision, 15.67° contact alignment, 1.07 mm max penetration, 0.00% frames with penetration $\mathcal{W}_\text{interaction}(i,j)=\text{clamp}\left(1-\frac{\mathcal{M_\text{dist}(i,j)-d_\text{min}}{d_\text{max}-d_\text{min}}\right), \qquad \mathcal{W}_\text{floor}(i)=\text{clamp}\left(1-\frac{\mathcal{M}_\text{height}(i)-h_\text{min}}{h_\text{max}-h_\text{min}}\right)$9 mm, and 4.70 ms/frame, outperforming the listed baselines on contact fidelity while remaining fast enough for real-time retargeting (Wu et al., 15 Jun 2026). On downstream RL, it reports 84.4% success on Ho-cap and 87.5% success on MoCap Pen-Spin, with the paper highlighting a 40.6 percentage-point improvement on Pen-Spin over the best existing baseline according to its statement (Wu et al., 15 Jun 2026). REGRIND evaluates on scissors and screwdriver tasks with LEAP and WUJI hands, showing that interaction-preserving retargeting yields near-perfect simulation success on three of four task-hand combinations and substantially better hardware transfer than the listed baselines (Feng et al., 13 Jul 2026).

For humanoid robot retargeting, Human2Humanoid evaluates downstream controllability with Success Rate and Tracking Error, and physical feasibility with Foot Skating and Ground Penetration. Averaged across evaluation motions, it reports SR $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 0, TE $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 1, FS $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 2, GP $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 3 cm, outperforming the listed baselines on average tracking error and average ground penetration while reducing foot skating relative to GMR and Unitree Retarget (Huang et al., 2 Jun 2026). OmniRetarget evaluates penetration duration, maximum penetration depth, foot-skating duration, maximum skating velocity, contact preservation, and downstream RL success for robot-object, terrain, and locomotion tasks. On OMOMO robot-object interaction, it reports $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 4 penetration duration, $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 5 cm max depth, $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 6 foot-skating duration, $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 7 maximum skating velocity, $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 8 contact preservation, and $\mathbf{d}^{t,i,j} = \mathbf{t}_i^{-1}\left(\mathbf{p}_j^t - \mathbf{p}_i^t\right),$ 9 RL success, improving over PHC, GMR, and VideoMimic on feasibility while maintaining strong contact preservation (Yang et al., 30 Sep 2025). DynaRetarget evaluates object-centric success after dynamic refinement and reports 74.6% success versus 37.9% for SPIDER under its object tracking thresholds, with smoother motions and better downstream policy success than raw OmniRetarget references (Dhedin et al., 6 Feb 2026).

For skinned-character retargeting, ReConForM evaluates jerk, self-penetration, floor penetration, grounded-feet detection, and foot-sliding detection. Its default $\text{normalized displacement} = \frac{\operatorname{FK}_k(q,t)-\operatorname{FK}_k(q^{\mathit{Tpose}})}{S}.$ 0 model reports jerk mean 212.8, self-penetration mean $\text{normalized displacement} = \frac{\operatorname{FK}_k(q,t)-\operatorname{FK}_k(q^{\mathit{Tpose}})}{S}.$ 1, floor penetration mean $\text{normalized displacement} = \frac{\operatorname{FK}_k(q,t)-\operatorname{FK}_k(q^{\mathit{Tpose}})}{S}.$ 2, grounded feet F1 0.925, and foot sliding F1 0.835, and the paper states a 70% reduction in floor penetration compared to the state of the art (Cheynel et al., 28 Feb 2025). Spatially adaptive anchor retargeting reports Pen $\text{normalized displacement} = \frac{\operatorname{FK}_k(q,t)-\operatorname{FK}_k(q^{\mathit{Tpose}})}{S}.$ 3, Prec 0.415, Rec 0.350, and Acc 0.948 overall, improving over MotionBuilder, SAME, R $\text{normalized displacement} = \frac{\operatorname{FK}_k(q,t)-\operatorname{FK}_k(q^{\mathit{Tpose}})}{S}.$ 4ET, and MeshRet on the listed interaction and penetration metrics (Choi et al., 19 May 2026). MeshRet reports Contact Error 0.772 on Mixamo+ and 0.284 on ScanRet, with Penetration 3.45% and 1.59% respectively, outperforming PMNet, SAN, and R $\text{normalized displacement} = \frac{\operatorname{FK}_k(q,t)-\operatorname{FK}_k(q^{\mathit{Tpose}})}{S}.$ 5ET on those measures (Ye et al., 2024). Contact-aware retargeting of skinned motion reports the best interpenetration score, vertex contact MSE, foot-contact accuracy, and global position MSE among its listed baselines, with encoder-space optimization clearly improving over its own IK-based variant (Villegas et al., 2021).

For multi-character interaction, the evidence is primarily qualitative. Simulation and retargeting of complex multi-character interactions shows that joint-wise imitation fails to preserve interaction geometry under size change, whereas the interaction-graph reward preserves greetings, support, Salsa grasping and support, and box carrying/throwing semantics across scaled humanoids and robot embodiments (Zhang et al., 2023). Physics-based sparse-input avatar retargeting likewise offers ablations showing that contact rewards shape gait style, orientation rewards preserve head-motion semantics, and asymmetric critic design is “essential for sparse real time retargeting” under difficult motions (Reda et al., 2023).

6. Limitations, fault lines, and emerging directions

A major fault line in the literature is between semantic interaction preservation and physical feasibility. Kinematic methods often preserve fine surface relations or contact regions well but may lack force, friction, and execution guarantees (Lakshmipathy et al., 2024, Cheynel et al., 28 Feb 2025). Physics-based methods preserve balance, contact plausibility, and dynamic realizability, but may be limited to coarser contact semantics such as stance timing, object trajectory, or gross hand-object geometry rather than exact contact manifolds (Reda et al., 2023, Dhedin et al., 6 Feb 2026). This suggests that no single representation currently covers dexterous contact topology, whole-body physics, and broad morphology variation equally well.

Another persistent limitation is scope of interaction. Several methods preserve only a subset of interactions: sparse-input physics retargeting preserves self-motion and environment contact interactions, especially foot-ground contacts and balance under gravity, but has no explicit object-contact model (Reda et al., 2023). Human2Humanoid focuses almost entirely on foot-ground interaction and support-phase structure rather than hand-object contact or scene-aware manipulation (Huang et al., 2 Jun 2026). AdaMorph preserves semantics, timing, and orientation evolution but does not explicitly preserve contacts, end-effector-object relations, or environment constraints (Zhang et al., 12 Jan 2026). Even strong interaction-mesh methods may not yet preserve fine dexterous contacts unless the representation is sufficiently dense and object-aware (Yang et al., 30 Sep 2025, Feng et al., 13 Jul 2026).

A third limitation concerns correspondence and embodiment mismatch. Many methods still assume shared topology or at least semantic joint or anchor correspondences. Adaptive-anchor methods explicitly target the failure of static correspondences under exaggerated body shape, but still assume shared skeleton topology at the level of joint-pooled features (Choi et al., 19 May 2026). MeshRet requires semantically consistent sensors and does not support missing limbs (Ye et al., 2024). Human2Humanoid avoids dense joint correspondence through skeleton-aware graph convolution and shared end-effectors, but remains limited to humanoid robots and support-contact preservation (Huang et al., 2 Jun 2026). Physics-based sparse-input retargeting uses manually designed sparse correspondences from human joints to conceptually similar creature joints and does not solve arbitrary topology mismatch in a principled way (Reda et al., 2023).

A fourth limitation is temporal and contact-horizon reasoning. Some systems are fundamentally framewise and rely only on local temporal regularization, which can weaken long-horizon contact progression or mode continuity (Lakshmipathy et al., 2024, Cheynel et al., 28 Feb 2025). DynaRetarget’s incremental-horizon SBTO is one response to this issue, specifically for long-horizon humanoid-object tasks (Dhedin et al., 6 Feb 2026). This suggests that future progress may continue to combine interaction-preserving representations with trajectory-level optimization, physics rollout, and downstream policy learning rather than relying on framewise retargeting alone.

Finally, the literature indicates an emerging synthesis. One direction preserves semantics of interaction rather than literal motion: root progression, head/root orientation, stance timing, end-effector-relative behavior, local hand-object topology, or multi-agent landmark relations (Reda et al., 2023, Huang et al., 2 Jun 2026, Zhang et al., 2023). Another direction makes these semantics executable through hard constraints, simulation, or residual RL (Yang et al., 30 Sep 2025, Dhedin et al., 6 Feb 2026, Feng et al., 13 Jul 2026). A plausible implication is that the strongest future systems will combine adaptive geometry-aware correspondences, interaction-centric representations, and physics-consistent refinement, so that interaction semantics remain intact while the target embodiment is allowed to realize them differently.