Advanced Motion Retargeting Techniques

• Advanced Motion Retargeting is a set of techniques that transfer motion sequences across characters with differing kinematics while preserving semantic intent and physical plausibility.
• The methods leverage skeletal-agnostic representations, transformer architectures, and unsupervised cycle-consistency to handle unpaired data and topology variations.
• These approaches enable real-time, physically feasible motion retargeting in animation, robotics, and embodied AI, reducing artifacts and ensuring reliable performance.

Advanced motion retargeting refers to the collection of computational methods, models, and optimization procedures that enable the transfer of motion from a source character (human, robot, or synthetic agent) to a target with different kinematic or morphological structure, character geometry, or embodiment. This task is central to animation, simulation, robotics, and embodied AI. Recent advances encompass skeleton-agnostic factorization, unsupervised correspondence learning, contact- and penetration-aware optimization, and unified transformer-based architectures enabling cross-morphology, cross-domain transfer.

1. Core Principles and Problem Formulation

Motion retargeting aims to map an input motion sequence from a source skeleton $\mathcal{S}_A$ (with $N_A$ joints, bone structure, and mesh) and motion $\mathcal{M}_A$ to a target skeleton $\mathcal{S}_B$ with different numbers of joints, kinematics, or geometry, producing a motion $\mathcal{M}_B$ that matches the semantic intent of the original, satisfies the target's physical and morphological constraints, and preserves geometric plausibility and temporal consistency.

Formally, given skeletal graphs $\mathcal{S}_A=(V_A,E_A)$ and $\mathcal{S}_B=(V_B,E_B)$, and corresponding motion sequences (typically joint rotations and global root trajectories), the retargeting function $f$ must generate:

$$\mathcal{M}_B = f(\mathcal{M}_A, \mathcal{S}_A, \mathcal{S}_B)$$

subject to:

• Semantic equivalence of the resulting movement (preserved style, function, intent)
• Kinematic and physical feasibility on the target
• Realistic and artifact-free geometry (no interpenetration, valid contact)

Key challenges arise from unpaired domains, variable topology, and the need for generalization to unseen skeletons, morphologies, or tasks.

2. Skeleton-Agnostic Representations and Disentanglement

A central advance is the development of skeletal-agnostic motion representations, where motion, structure, and view (camera) are explicitly disentangled. Canonical examples include:

• MoCaNet (Zhu et al., 2021): Decomposes a 2D skeleton sequence $x\in\mathbb{R}^{2N\times T}$ into three independent codes:
  • Motion $m$ (temporal dynamics, $\mathbb{R}^{M\times C_m}$)
  • Structure $s$ (bone-length ratios, $\mathbb{R}^{C_s}$)
  • View $v$ (per-frame orientation, $\mathbb{R}^{T\times6}$)
  • Canonicalization operations enforce invariance such that retargeting is done by swapping $m$ and $s$ between source and target, then decoding into 3D skeletons.
• HuMoT (Mourot et al., 2023): Implements a topology-agnostic transformer autoencoder, conditioning on "skeleton templates" to enable cross-topology transfer. The encoder takes joint-wise global-space positions with template embeddings, creating a latent code that can be decoded onto any skeleton with a specified template. Losses enforce both reconstruction and temporal bone-length consistency, allowing the direct retargeting of style and movement to previously unseen skeletons.
• Part-based and Body-part Pooling Methods: Partitioning the skeleton into meaningful parts (e.g., torso, limbs, head) guides attention and pooling, allowing the model to extract shared representations across topologies (Liu et al., 12 Jan 2026).

These methods use explicit architectural and loss design to ensure motion information is represented independently of bone graph topology or body morphology.

3. Unsupervised and Reversible Cross-Morphology Retargeting

Advanced retargeting frameworks now operate without paired motion data, leveraging unsupervised, cycle-consistent, and flow-matching strategies for aligning the motion spaces of diverse characters:

• MoReFlow (Kim et al., 29 Sep 2025) constructs discrete, tokenized VQ-VAE latent spaces for each character, then employs unsupervised flow matching to learn mappings between these latent distributions. The flow-matching vector field is trained using optimal transport between motion tokens, with classifier-free guidance enabling conditional control and bidirectionality. The result is a fully reversible framework supporting multi-character, multi-morphology transfer without paired data.
• ACE (Li et al., 2023) pretrains a motion prior on the target character and adversarially learns a correspondence embedding from source-human to target-robot/creature motions. Feature-consistency losses align high-level statistics (root, end-effector motion), while adversarial training ensures domain realism.
• TransMoMo (Yang et al., 2020), Learning Character-Agnostic Motion in 2D (Aberman et al., 2019), and Skeleton-Aware Networks (Aberman et al., 2020) all implement disentangled/unsupervised factorizations and cycle-consistency or adversarial losses to allow cross-structural retargeting in 2D or 3D, tolerating substantial domain variation and eliminating the need for motion pairing.

Empirical results show state-of-the-art performance in both quantitative metrics (e.g., mean per-joint error, FID, contact accuracy) and user studies.

4. Constraint-Based and Physically Feasible Motion Mapping

Physical plausibility requires enforcing kinematic, dynamic, and contact constraints during retargeting:

• Contact-aware and Penetration-Aware Methods: Methods such as Contact-Aware Retargeting (Villegas et al., 2021), STaR (Yang et al., 9 Apr 2025), and R²ET (Zhang et al., 2023) include explicit loss terms or optimization steps to preserve self-contact, prevent interpenetration, and maintain physical realism. Typical techniques involve mesh-level energy terms (distance fields, SDF, vertex-pair losses), foot-ground contact preservation, and encoder-space refinement.
• Multi-Contact Whole-Body Optimization: Retargeting for high-DoF robots and loco-manipulation requires solving for joint positions and contact wrenches under equilibrium constraints. Sequential Quadratic Programming (SQP) (Rouxel et al., 2022) and related optimization approaches are used to solve, at each control step, for feasible joint/force trajectories that respect all kinematic, friction, and equilibrium boundaries at kilohertz rates—ensuring real-time retargeting for robot operation.
• Curriculum and Physics Losses: Transformer-based models for robots (e.g., AdaMorph (Zhang et al., 12 Jan 2026)) integrate curriculum-based trajectories, orientation geodesics, and physical trajectory consistency losses to ensure that generated motions respect both orientation and trajectory constraints across highly heterogeneous morphologies.

These constraints are critical in teleoperation, animation, and robotics to prevent physically impossible postures or contacts.

5. Real-Time Semantic and Contact Preservation

Recent methods model contact explicitly and adapt to challenging semantics such as multi-character scenes, complex contact sequences, and environmental adaptation:

• ReConForM (Cheynel et al., 28 Feb 2025): Proposes a key-vertex, descriptor-based optimization framework where a sparse set of mesh-anchored semantic vertices are matched and descriptors (distance, penetration, height, sliding) are adaptively weighted in the optimization objective. Proximity-based weights focus computation on relevant features (e.g., when a hand approaches a surface). Extensions enable multi-character and uneven terrain via adapted descriptors.
• STaR (Yang et al., 9 Apr 2025) combines spatial modules (limb penetration loss, mesh-based attention) with temporal transformers (multi-level trajectory consistency) for smooth and plausible retargeted motion. The result is a reduction of both geometric and temporal artifacts.
• Contact-Aware Retargeting (Villegas et al., 2021) integrates geometry-conditioned RNNs and per-frame optimization in latent space to more accurately enforce contacts and avoid mesh penetration at runtime.

These systems are designed to operate at real-time rates and preserve semantic features across retargeted character populations.

6. Domain-Generalization, Adaptation, and Robotics Applications

Generalization across morphologies is a core challenge for deployment in animation, simulation, and real robots.

• Unified Models across Robots: Architectures such as AdaMorph (Zhang et al., 12 Jan 2026) and MR HuBo (Figuera et al., 2024) provide mechanisms—conditional transformers, dual-pathway prompting, reverse pairing—for transferring motion across robots with arbitrarily differing kinematics. MR HuBo reverses the pairing process, mapping from robot space to human, leveraging a human-body prior (VPoser) for high-quality paired data.
• Sim-to-Real and Safety Guarantees: Approaches employing shared latent embeddings and projection-invariant mappings (Choi et al., 2021), or reinforcement learning with cyclic consistency (Kim et al., 2019), are deployed to enable safe and feasible real-robot retargeting. Explicit nonparametric projections enforce joint limit, self-collision, and physical validity.
• World-Coordinate Recovery for Embodied Intelligence: Lightweight pipelines recover metrically scaled, world-coordinate human motion from monocular input, applying smoothing, contact probability models, and two-stage IK to yield temporally consistent and robot-ready movement (Tu et al., 25 Dec 2025).

The resulting frameworks support teleoperation, embodied learning, multi-modal avatars, and physical robot control in both simulation and real environments.

7. Summary Table of Advanced Motion Retargeting Frameworks

Method / Paper	Key Advance	Domain / Morphology	Constraints Modeled	Supervision
MoCaNet (Zhu et al., 2021)	Unsupervised canonicalization, disentangled latent (M,B,V)	2D→3D, arbitrary shape, in-the-wild	Implicit via loss	Unsupervised
HuMoT (Mourot et al., 2023)	Topology-agnostic transformer, template conditioning	Cross-topology, unseen skeletons	Bone consistency	Unsupervised
MoReFlow (Kim et al., 29 Sep 2025)	Unsupervised, reversible flow matching in VQ-VAE latent	Multi-morph, bidirectional	Style, task alignment	Unsupervised
ACE (Li et al., 2023)	GAN + motion latent, feature loss	Human→robot/creature	Adversarial + correspondence	Unpaired
ReConForM (Cheynel et al., 28 Feb 2025)	Key-vertex semantic optimization, adaptive weighting	Mesh, arbitrary topology	Contact, penetration	Unsupervised
AdaMorph (Zhang et al., 12 Jan 2026)	Unified transformer, AdaLN, robot/human dual prompt	12 humanoid robots, unseen morph	Kinematic+traj, physics consistency	Supervised
Contact-Aware Retargeting (Villegas et al., 2021)	Self-contact, interpenetration, geometry-RNN and ESO	Mesh-based, arbitrary shapes	Contact, penetration	Unsupervised
MR HuBo (Figuera et al., 2024)	Paired-data via reverse mapping, VPoser prior	Humanoid robots	Human-like prior, contact	Supervised
STaR (Yang et al., 9 Apr 2025)	Spatial+temporal transformer, limb penetration loss	Human, stylized, real characters	Penetration, temporal smooth	Supervised

• "MoCaNet: Motion Retargeting in-the-wild via Canonicalization Networks" (Zhu et al., 2021)
• "HuMoT: Human Motion Representation using Topology-Agnostic Transformers for Character Animation Retargeting" (Mourot et al., 2023)
• "MoReFlow: Motion Retargeting Learning through Unsupervised Flow Matching" (Kim et al., 29 Sep 2025)
• "ACE: Adversarial Correspondence Embedding for Cross Morphology Motion Retargeting from Human to Nonhuman Characters" (Li et al., 2023)
• "PALUM: Part-based Attention Learning for Unified Motion Retargeting" (Liu et al., 12 Jan 2026)
• "AdaMorph: Unified Motion Retargeting via Embodiment-Aware Adaptive Transformers" (Zhang et al., 12 Jan 2026)
• "Contact-Aware Retargeting of Skinned Motion" (Villegas et al., 2021)
• "STaR: Seamless Spatial-Temporal Aware Motion Retargeting with Penetration and Consistency Constraints" (Yang et al., 9 Apr 2025)
• "ReConForM : Real-time Contact-aware Motion Retargeting for more Diverse Character Morphologies" (Cheynel et al., 28 Feb 2025)
• "Redefining Data Pairing for Motion Retargeting Leveraging a Human Body Prior" (Figuera et al., 2024)
• "World-Coordinate Human Motion Retargeting via SAM 3D Body" (Tu et al., 25 Dec 2025)
• "Self-Supervised Motion Retargeting with Safety Guarantee" (Choi et al., 2021)
• "C-3PO: Cyclic-Three-Phase Optimization for Human-Robot Motion Retargeting based on Reinforcement Learning" (Kim et al., 2019)
• "Learning Character-Agnostic Motion for Motion Retargeting in 2D" (Aberman et al., 2019)
• "Skeleton-Aware Networks for Deep Motion Retargeting" (Aberman et al., 2020)
• "Skinned Motion Retargeting with Residual Perception of Motion Semantics & Geometry" (Zhang et al., 2023)