Physics-Aware Motion Style Regularization

• The paper introduces PMSR to enforce physical plausibility in personalized motion by preventing unnatural limb stretching and preserving full-body structure.
• It employs coupled constraints of dynamic bone stability and body connectivity to ensure smooth bone-length evolution and coherent skeletal topology.
• Empirical evaluations within PersonaAnimator demonstrate that PMSR enhances style consistency and physical realism compared to baseline transfer methods.

to=arxiv_search 山大发cript _日本毛片免费视频观看 code: {"query":"(Qian et al., 27 Aug 2025) PersonaAnimator: Personalized Motion Transfer from Unconstrained Videos", "max_results": 5} to=arxiv_search мәҗбурcript 买天天中彩票 code: {"query":"(Peng et al., 2021) AMP: Adversarial Motion Priors for Stylized Physics-Based Character Control", "max_results": 5} to=search_arxiv 北京赛车能cript 天天中彩票一等奖 code: {"query":"PersonaAnimator Personalized Motion Transfer from Unconstrained Videos", "max_results": 10} Physics-aware motion style regularization denotes a family of mechanisms that constrain motion generation, transfer, retargeting, or recovery so that stylistic fidelity is coupled to physical plausibility rather than visual resemblance alone. In PersonaAnimator, the term refers specifically to Physics-aware Motion Style Regularization (PMSR), introduced for Video-to-Video Motion Personalization to prevent unnatural limb stretching, loss of structural cohesion, joint collapse or intersection, and distortion of full-body skeleton topology while learning personalized motion directly from unconstrained videos (Qian et al., 27 Aug 2025). In adjacent work, analogous roles are played by adversarial motion priors, decoded-motion contact and smoothness penalties, feasibility constraints, probabilistic dynamics residuals, and simulation-coupled retargeting objectives, all of which regularize style by restricting motion to a physically acceptable manifold (Peng et al., 2021).

1. Motivation and failure modes

The immediate motivation for PMSR is that visually convincing style transfer can still yield physically implausible motion. PersonaAnimator identifies several characteristic failure modes: limbs may stretch unnaturally, connected joints may lose structural cohesion, unrelated joints may become too close or intersect, and the full-body skeleton topology may be distorted. The underlying claim is not merely that style transfer can be inaccurate, but that style transfer can be successful in appearance while still violating basic geometric and temporal regularities of articulated motion (Qian et al., 27 Aug 2025).

Within the same problem setting, earlier motion style transfer methods are described as relying heavily on motion capture data, while pose-guided character motion transfer methods are described as merely replicating motion without learning its style characteristics. PersonaAnimator further notes that prior regularization strategies, including those used in MoST and Arbitrary Motion Style Transfer, partially mitigate implausibility with velocity and foot-contact constraints, yet remain insufficient because they do not impose explicit constraints on the full-body skeleton and do not model repulsion between non-connected joints. This suggests that local dynamic constraints and end-effector heuristics alone are inadequate when the objective is global physical plausibility under personalized style transfer from unconstrained video (Qian et al., 27 Aug 2025).

2. Placement within the PersonaAnimator framework

PersonaAnimator formulates Video-to-Video Motion Personalization as a two-stage pipeline. The first stage, Semantic-Aware Personalized Motion Transfer (SA-PMT), learns a personalized pose representation from a content motion video $\bm{V}_c$, a style motion video $\bm{V}_s$, and semantic content labels. The second stage performs personalized motion generation by conditioning a video diffusion model $\epsilon_\theta$ on a reference image $\bm{I}^r$, appearance latent $\bm{f}_a^r$, VAE latent $\bm{f}_e^r$, personalized motion feature $\bm{Z}_{pm}$, and noise latent $\bm{Z}_t$. PMSR is applied during training as a regularizer on the generated personalized pose sequence $\bm{P}_s^m$, so that learned personalization remains physically feasible rather than only stylistically expressive (Qian et al., 27 Aug 2025).

The same framework introduces PersonaVid, described as the first video-based personalized motion dataset, containing 20 motion content categories and 120 motion style categories. PMSR is optimized jointly with style consistency and reconstruction-oriented objectives. The total loss is given as

$$\mathcal{L}_{\text{total}}=\mathcal{L}_{\text{stability}}+\mathcal{L}_{\text{conn}}+\mathcal{L}_{sc}+\mathcal{L}_d,$$

where $\bm{V}_s$0 is the dynamic bone stability term, $\bm{V}_s$1 is the body connectivity term, $\bm{V}_s$2 is the style consistency loss from Aberman et al., and $\bm{V}_s$3 is the MSE loss for appearance alignment and motion modeling. In the paper’s own role division, SA-PMT determines “what style to transfer,” whereas PMSR determines “how to keep it physically valid” (Qian et al., 27 Aug 2025).

3. Mathematical structure of PMSR

PMSR consists of two coupled components: dynamic bone stability constraints and body connectivity constraints. The first component penalizes acceleration in bone-length changes over time. Let $\bm{V}_s$4 denote the 2D coordinate of joint $\bm{V}_s$5 in frame $\bm{V}_s$6, let $\bm{V}_s$7 denote the set of bones, and let $\bm{V}_s$8 denote the length of bone $\bm{V}_s$9 at frame $\epsilon_\theta$0. Bone length is defined by

$\epsilon_\theta$1

The second-order temporal difference is

$\epsilon_\theta$2

and the stability loss is written as

$\epsilon_\theta$3

Its stated purpose is to encourage bone lengths to vary smoothly and to prevent abrupt stretching or compression (Qian et al., 27 Aug 2025).

The second component encodes skeleton topology by combining attraction for connected joints and repulsion for unconnected joints. An adjacency matrix

$\epsilon_\theta$4

is defined so that $\epsilon_\theta$5 if joints $\epsilon_\theta$6 and $\epsilon_\theta$7 should be connected and $\epsilon_\theta$8 otherwise. For each frame $\epsilon_\theta$9, a distance matrix is constructed as

$\bm{I}^r$0

The connected-joint attraction term is

$\bm{I}^r$1

and the unconnected-joint repulsion term is

$\bm{I}^r$2

with minimum distance threshold

$\bm{I}^r$3

The combined connectivity loss is

$\bm{I}^r$4

In the paper’s interpretation, this is the principal novelty over earlier style regularizers because it explicitly encodes skeleton topology and a non-penetration-like separation prior rather than relying only on motion smoothness (Qian et al., 27 Aug 2025).

4. Physical effects, training role, and empirical behavior

The physical properties encouraged by PMSR are stated directly. Dynamic bone stability requires bone lengths to evolve smoothly over time and avoids sudden stretching or shrinking. Body connectivity preservation requires connected joints to remain sufficiently close, disconnected joints not to collapse into one another, and the overall skeleton structure to remain coherent. Together, these mechanisms are intended to reduce limb distortion, broken skeleton topology, and implausible overlap or collapse of body parts. Because PersonaAnimator learns from unconstrained videos rather than motion-capture data, the paper emphasizes that noisy pose estimates and uncontrolled camera or viewpoint variation can otherwise produce implausible pose sequences, which makes regularization structurally important rather than ancillary (Qian et al., 27 Aug 2025).

The reported evidence is primarily qualitative. In comparative results, PersonaAnimator is said to capture style traits such as a Taichi-style lunge pose and both hands sweeping left to right, while baseline methods may miss hand motion, produce only partial body motion, or generate distorted poses. The paper explicitly attributes these results to the combination of the SA-PMT module and PMSR: “These results stem from our SA-PMT module, which semantically guides the learning of motion features, and the PMSR mechanism, which enforces physical plausibility and prevents motion distortion.” On an unseen skating style, PersonaAnimator is reported to capture personalized motion while avoiding mechanical copying, whereas baselines produce unnatural poses or fail to learn personalization. Quantitatively, the excerpt does not provide a PMSR-only ablation table, but it states that the full PersonaAnimator framework achieves state-of-the-art results in both qualitative and quantitative comparisons on PersonaVid and mainstream human animation datasets. The overall model is trained and evaluated on PersonaVid, trained on NVIDIA A800, uses RMSprop, and uses a learning rate of $\bm{I}^r$5 (Qian et al., 27 Aug 2025).

5. Related formulations in control, diffusion, retargeting, and recovery

In physics-based character control, AMP provides a distinct but closely related formulation. Rather than regularizing a generated pose sequence with explicit skeleton constraints, AMP trains an adversarial motion prior that converts a discriminator over state transitions into a style reward, and the policy maximizes a weighted sum of task reward and style reward in a physics simulator. With

$\bm{I}^r$6

and with $\bm{I}^r$7 in the reported experiments, the prior specifies “how to do it” while the task reward specifies “what to do.” The motion dataset is treated as an unstructured style distribution rather than as a trajectory to be exactly tracked, and the policy must still obey contacts, balance, and task constraints under simulation (Peng et al., 2021).

In robot-oriented style transfer, the bionic generation-to-control framework for humanoid robots applies physics-aware regularization during diffusion training on the decoded clean-motion estimate. Two penalties are imposed: a contact-consistency loss suppressing stance-foot sliding and ground penetration, and a temporal-smoothness loss matching predicted and reference velocity and acceleration structure. These terms are computed from decoded motion, not latent noise, and only during training. In the reported ablation, using both regularizers gives FSF $\bm{I}^r$8, $\bm{I}^r$9 $\bm{f}_a^r$0, and $\bm{f}_a^r$1 $\bm{f}_a^r$2, compared with FSF $\bm{f}_a^r$3, $\bm{f}_a^r$4 $\bm{f}_a^r$5, and $\bm{f}_a^r$6 $\bm{f}_a^r$7 without both regularizers. The same paper reports a $\bm{f}_a^r$8 success rate over $\bm{f}_a^r$9 reported real-robot trials and frames the regularization as improving the reference before control rather than guaranteeing perfect hardware behavior (Huang et al., 2 Jun 2026).

In cross-morphology retargeting, Human2Humanoid imposes explicit feasibility constraints in a CycleGAN-based architecture. The source-conditioned foot contact constraint penalizes target foot velocity when the source motion indicates contact, the foot-height constraint discourages hovering or floating during stance, and the joint-limit term penalizes excursions outside allowable mechanical ranges. In its ablation, removing $\bm{f}_e^r$0 and $\bm{f}_e^r$1 drops success rate from $\bm{f}_e^r$2 to $\bm{f}_e^r$3, increases tracking error from $\bm{f}_e^r$4 to $\bm{f}_e^r$5, increases foot skating from $\bm{f}_e^r$6 to $\bm{f}_e^r$7, and increases ground penetration from $\bm{f}_e^r$8 cm to $\bm{f}_e^r$9 cm. The full system reports average SR $\bm{Z}_{pm}$0, TE $\bm{Z}_{pm}$1, FS $\bm{Z}_{pm}$2, and GP $\bm{Z}_{pm}$3 cm on Unitree G1 (Huang et al., 2 Jun 2026).

In hand motion recovery, PAD-Hand uses a probabilistic dynamics formulation rather than deterministic topology or contact penalties. It models Euler–Lagrange residuals as virtual observables,

$\bm{Z}_{pm}$4

and defines a physics likelihood term

$\bm{Z}_{pm}$5

A last-layer Laplace approximation produces per-joint, per-time variances that serve as interpretable indicators of where physical consistency weakens. This formulation differs from zero-residual enforcement by explicitly acknowledging noise, model mismatch, and uncertainty in inferred inertial properties (Ismayilzada et al., 27 Mar 2026).

In retargeting by reinforcement learning, ReActor embeds style preservation inside a bilevel optimization problem,

$\bm{Z}_{pm}$6

where the upper level preserves source motion structure and the lower level trains a tracking policy with both tracking and regularization rewards. The regularization terms penalize joint torque, joint acceleration, action rate, action acceleration, root force, and root torque, thereby suppressing physically implausible control strategies while keeping the retargeted motion close to the source’s kinematic signature (Müller et al., 7 May 2026).

6. Conceptual boundaries, misconceptions, and limitations

A recurrent misconception is that physics-aware motion style regularization is equivalent to exact tracking of a reference clip. AMP directly rejects this equivalence: the discriminator is trained on state transitions from a motion corpus, not on phase-synchronized target frames, and the policy is rewarded for producing transitions that resemble the style dataset rather than for minimizing error to a single reference trajectory. This makes style regularization distributional rather than frame-locked, and it explains why automatic composition and interpolation of behaviors can emerge without explicit clip selection or motion graphs (Peng et al., 2021).

A second misconception is that “physics-aware” necessarily implies a full rigid-body guarantee. Several of the cited works state the opposite in precise terms. Human2Humanoid characterizes its feasibility terms as inexpensive kinematic constraints rather than a full dynamics-aware contact or balance controller. The humanoid diffusion framework states that its regularizers reduce artifacts but do not guarantee exact executability after retargeting. PAD-Hand argues that driving Euler–Lagrange residuals to zero is often too strict because the motion input is noisy, the physics model is approximate, inertial parameters are estimated, and contacts may be unmodeled. ReActor similarly notes that some motions are physically impossible for the target embodiment and that the tradeoff between style and feasibility becomes more delicate as morphology diverges (Huang et al., 2 Jun 2026, Huang et al., 2 Jun 2026, Ismayilzada et al., 27 Mar 2026, Müller et al., 7 May 2026).

Within this landscape, PMSR occupies a specific point: it is not a simulator-based controller, not a hard contact solver, and not a probabilistic dynamics model. It is a topology-aware regularizer for personalized motion learned from unconstrained video, constructed from dynamic bone stability and connected/unconnected joint relations. This suggests that physics-aware motion style regularization is best understood not as a single algorithmic template but as a design principle: style transfer remains the primary objective, but the permissible style manifold is narrowed by priors that encode articulated structure, contact semantics, mechanical feasibility, or explicit dynamics, depending on the application domain (Qian et al., 27 Aug 2025).