PhysDynPose: Physics-Informed Pose Estimation

• PhysDynPose is a family of methods that integrate 2D/3D pose estimations with biomechanical and physics-based constraints to ensure dynamic consistency.
• The framework employs techniques like differentiable inverse dynamics, closed-loop filtering, and contact modeling to mitigate noise amplification and enforce rigid-body mechanics.
• It enables real-time and scene-aware optimization for applications in sports science, rehabilitation, animation, and physiotherapy by moving beyond purely geometric reconstructions.

Searching arXiv for the cited PhysDynPose-related papers to ground the article. arXiv search query: PhysDynPose PhysDynPose denotes a line of physics-informed human pose estimation systems that augment monocular video kinematics with dynamic consistency, inverse dynamics, contact reasoning, scene geometry, or anatomical priors. In the supplied literature, the name is used for a real-time post-processing algorithm that fuses BlazePose 2D and 3D landmarks with biomechanical constraints (Leuthold et al., 7 Dec 2025), for a physics-based method that combines 4DHumans, DROID-SLAM, and a scene-aware optimizer for a single moving RGB camera (Aytekin et al., 23 Jul 2025), and for a differentiable inverse-dynamics design direction built around SMPL-Dynamics and torque plausibility (Kim et al., 23 May 2026). Across these formulations, the common objective is to move beyond purely geometric reconstruction toward estimates that remain compatible with rigid-body mechanics, support conditions, and biomechanical regularity.

1. Scope and conceptual lineage

The label does not identify a single canonical architecture. In the supplied literature, it spans at least three closely related constructions: an online backend refinement pipeline operating on BlazePose outputs (Leuthold et al., 7 Dec 2025), a scene-aware physics optimizer for moving-camera capture (Aytekin et al., 23 Jul 2025), and a framework direction in which inverse dynamics is embedded into training or test-time refinement through a differentiable layer (Kim et al., 23 May 2026). This suggests PhysDynPose is best understood as a family of methods rather than a standardized implementation.

Its methodological lineage includes several adjacent systems that instantiate the same physics-informed objective under different names. “Physics-Guided Human Motion Capture with Pose Probability Modeling” models pose uncertainty with a latent Gaussian, performs reverse diffusion, and injects physics-based reprojection gradients from MuJoCo tracking into denoising (Ju et al., 2023). “D&D: Learning Human Dynamics from Dynamic Camera” addresses non-inertial camera motion through Inertial Force Control, Probabilistic Contact Torque, and an Attentive PD Controller, while remaining entirely neural-based and avoiding offline optimization or simulation in physics engines (Li et al., 2022). “Contact and Human Dynamics from Monocular Video” couples learned contact timing with centroidal-dynamics trajectory optimization (Rempe et al., 2020). “3D Human Pose Estimation via Intuitive Physics” introduces differentiable support and balance surrogates based on pressure, Center of Pressure, and Center of Mass (Tripathi et al., 2023). “Optimal-state Dynamics Estimation for Physics-based Human Motion Capture from Videos” adds an online neural Kalman-filter loop around a physics simulator and a meta-PD controller (Le et al., 2024).

A recurring theme across these works is dissatisfaction with kinematics-only pipelines. The motivating applications are consistently those in which joint positions are not sufficient: sports science, rehabilitation, ergonomics, physics-based animation, automated physiotherapy, healthcare, and sports coaching all require estimates that remain mechanically plausible under forces, torques, contact, and temporal smoothness (Kim et al., 23 May 2026, Leuthold et al., 7 Dec 2025).

2. Mechanical formulation and state representation

Most PhysDynPose formulations begin from generalized coordinates. In the inverse-dynamics setting specialized to the SMPL kinematic tree, the pose is written as $\theta \in \mathbb{R}^n$, and the joint torque vector $\tau \in \mathbb{R}^n$ satisfies

$$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$

where $M(\theta)$ is the joint-space mass matrix, $C(\theta,\dot\theta)\dot\theta$ contains Coriolis and centrifugal effects, and $G(\theta)$ is the gravity term (Kim et al., 23 May 2026). The same paper also gives a recursive Newton–Euler form over the tree, with per-joint transmitted force and torque accumulated from children.

Scene-aware and contact-aware formulations extend this to floating-base articulated dynamics. In the moving-camera setting, the generalized coordinates are written as $q=[r;\theta]\in\mathbb{R}^{6+J}$, and the rigid-body equation takes the form

$$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$

with contact Jacobian $J_c(q)$ and contact forces $\lambda$ (Aytekin et al., 23 Jul 2025). The accompanying constraints include unilateral contact,

$\tau \in \mathbb{R}^n$0

and the friction cone,

$\tau \in \mathbb{R}^n$1

In practice, this formulation is combined with image reprojection, physical residuals, contact penalties, and a statistical pose prior within a timestep-wise optimization.

Dynamic-camera formulations additionally model non-inertial effects. D&D introduces an inertial torque term

$\tau \in \mathbb{R}^n$2

which depends on linear camera acceleration $\tau \in \mathbb{R}^n$3 and angular velocity $\tau \in \mathbb{R}^n$4, and adds it to local-frame dynamics to explain motion in a non-inertial camera coordinate system (Li et al., 2022). This is a direct response to the fact that camera acceleration changes the apparent force balance in the local frame.

At a broader estimation-theoretic level, the variational background is provided by rigid-body pose estimation on $\tau \in \mathbb{R}^n$5 via the Lagrange–d’Alembert principle. That construction defines a kinetic-energy-like term, artificial potentials for rotational and translational error, and a Rayleigh-type dissipation term, yielding stable asymptotic convergence in the absence of measurement noise and a Lie-group variational integrator for discretization (Izadi et al., 2015). Although not human-specific, it provides a formal template for dynamics-consistent pose estimation under noisy measurements.

3. Noise amplification, filtering, and differentiable inverse dynamics

A central result in the recent PhysDynPose direction is that inverse dynamics is highly sensitive to pose noise. “How Noisy Poses Break Inverse Dynamics” reports three key findings: pose noise is amplified by approximately $\tau \in \mathbb{R}^n$6 when computing joint torques via numerical differentiation, proximal joints such as the spine and hips are up to $\tau \in \mathbb{R}^n$7 more sensitive to noise than distal joints such as wrists and hands, and low-pass filtering before differentiation substantially reduces this amplification (Kim et al., 23 May 2026).

The mechanism is explicit. If the angle estimate has zero-mean Gaussian noise $\tau \in \mathbb{R}^n$8, then numerical differentiation induces $\tau \in \mathbb{R}^n$9 and $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$0, so the torque amplification factor is of order

$$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$1

At $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$2 s, this yields $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$3, matching the reported empirical amplification of approximately $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$4 for $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$5 rad, with average torque error around $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$6 Nm versus pose error around $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$7 rad (Kim et al., 23 May 2026).

To analyze and mitigate this effect, the paper introduces SMPL-Dynamics, a fully differentiable inverse dynamics module for the SMPL body model that requires no external physics simulators. It accepts a temporal SMPL pose sequence and root translations, performs forward kinematics, applies a low-pass filter in pose space, computes finite-difference angular and linear velocities and accelerations, and then runs a leaf-to-root Recursive Newton–Euler pass. Because matrix logs, finite differences, and cross-products all have backward definitions, gradients $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$8 flow naturally through the entire module (Kim et al., 23 May 2026).

The recommended denoising step is a zero-phase forward–backward 4th-order Butterworth filter applied independently to each axis-angle channel. In the reported implementation, $$\tau(\theta,\dot\theta,\ddot\theta) = M(\theta)\,\ddot\theta + C(\theta,\dot\theta)\,\dot\theta + G(\theta),$$9 Hz yields approximately $M(\theta)$0 reduction in torque error while preserving natural motion up to approximately $M(\theta)$1 Hz harmonics (Kim et al., 23 May 2026). The same differentiability enables pose refinement through a torque-based objective,

$M(\theta)$2

with $M(\theta)$3 Nm, $M(\theta)$4, $M(\theta)$5, and $M(\theta)$6, optimized with Adam for $M(\theta)$7 iterations. On a walking sequence with $M(\theta)$8 rad added noise, the reported torque error drops from $M(\theta)$9 Nm to $C(\theta,\dot\theta)\dot\theta$0 Nm, a $C(\theta,\dot\theta)\dot\theta$1 reduction, while pose deviation changes from $C(\theta,\dot\theta)\dot\theta$2 rad to $C(\theta,\dot\theta)\dot\theta$3 rad; the abstract summarizes this as a $C(\theta,\dot\theta)\dot\theta$4 reduction in torque error with negligible change in pose (Kim et al., 23 May 2026).

4. Contact, support, and scene geometry

A major branch of PhysDynPose focuses on contact timing and scene consistency. In “Contact and Human Dynamics from Monocular Video,” contact is first estimated by a dedicated network operating on a 9-frame window of lower-body OpenPose detections. The network predicts four binary labels—left-toe, left-heel, right-toe, right-heel contact—and the resulting labels seed a centroidal-dynamics trajectory optimization over COM position, COM orientation, foot positions, contact forces, and contact-phase durations (Rempe et al., 2020). The optimization enforces Newton’s and Euler’s equations, no-slip constraints during stance, zero force during flight, a friction pyramid with $C(\theta,\dot\theta)\dot\theta$5, constant toe–heel distance, and a leg-length bound, while fitting the kinematic input.

The moving-camera PhysDynPose extends contact modeling to world-frame scene interaction. Its first stage uses 4DHumans for per-frame SMPL pose parameters and DROID-SLAM for per-frame camera poses; root translation is converted from camera to world coordinates through

$C(\theta,\dot\theta)\dot\theta$6

A floating-base humanoid is then initialized in PyBullet, and a constrained quadratic program solves for generalized accelerations, contact forces, and torques that track the reference pose, enforce Newton–Euler dynamics, respect friction cones and non-penetration against a scene height map, and apply root supervision from future frames to reduce drift (Aytekin et al., 23 Jul 2025). This design is specifically intended for non-flat scenes and moving cameras, where flat-ground assumptions and camera-static formulations break down.

An alternative route replaces explicit contact simulation with differentiable support surrogates. IPMAN defines a per-vertex pressure field

$C(\theta,\dot\theta)\dot\theta$7

projects it to the ground plane, computes the Center of Pressure as a weighted average of projected vertex locations, and estimates the Center of Mass by a part-mass construction rather than naive vertex averaging (Tripathi et al., 2023). The resulting stability loss,

$C(\theta,\dot\theta)\dot\theta$8

is coupled with a ground-contact loss that penalizes penetration and softly encourages contact. In this formulation, floor support becomes a differentiable prior that can be inserted into either regression or optimization.

These variants show that “contact” in PhysDynPose is not a single object. It may be a binary label over toes and heels, a force variable within a constrained optimizer, a probabilistic latent variable, or a differentiable pressure surrogate. What remains constant is the role of support as a primary regularizer of otherwise underconstrained monocular reconstruction.

5. Control loops, filtering, denoising, and anatomical optimization

Several PhysDynPose formulations are organized as explicit closed-loop estimators. OSDCap begins from a noisy kinematics estimate from TRACE and the last filtered state, predicts torques and contact forces with a learned meta-PD law, simulates forward dynamics with an inertia correction term, and then fuses the simulated prediction with the next kinematic observation through a learned Kalman gain produced by a GRU (Le et al., 2024). The update step,

$C(\theta,\dot\theta)\dot\theta$9

makes the filtered pose the “optimal state,” which is then reused as the controller target at the next timestep. The paper emphasizes that this online filtering step is crucial for balancing noisy observations and imperfect simulation.

D&D pursues a fully neural dynamics route. It predicts contact probabilities $G(\theta)$0, soft-samples contact indicators $G(\theta)$1 through a Gumbel–Softmax relaxation, and uses the resulting probabilistic contact torque

$G(\theta)$2

inside inertial forward dynamics (Li et al., 2022). In parallel, an Attentive PD Controller replaces the raw target pose by a temporally smoothed attention-weighted target, then predicts time-varying proportional and derivative gains together with an offset. The system remains end-to-end differentiable, does not invoke a black-box physics engine, and is designed for moving-camera videos.

Physics-guided diffusion introduces a probabilistic denoising interpretation. A GRU-based encoder maps image features to per-frame Gaussian parameters, a modified DDPM performs reverse denoising, and a MuJoCo character tracked by a fixed policy yields reprojection-loss gradients that are concatenated with image features and injected back as reverse-diffusion conditions (Ju et al., 2023). Here physics acts neither as a separate post-processor nor as a hard constraint layer, but as an iterative denoising guide.

The BlazePose-based PhysDynPose targets real-time posture estimation with explicit anatomical priors. At each frame it minimizes

$G(\theta)$3

where the terms enforce world-likeliness, line-of-sight consistency, single-bone ratio consistency, and multi-bone biomechanical constraints (Leuthold et al., 7 Dec 2025). Bone ratios are tracked by an adaptive Kalman filter with

$G(\theta)$4

so lower confidence increases the measurement covariance. The reported inference loop runs at 30 Hz, applies a 4th-order Butterworth low-pass filter with 10 Hz cutoff, updates bone-ratio priors from high-confidence measurements, initializes L-BFGS-B with the previous optimized solution, and outputs a refined 3D pose for downstream coaching or analysis (Leuthold et al., 7 Dec 2025).

Taken together, these systems show four distinct integration patterns: hard-constrained optimization, differentiable inverse-dynamics loss layers, closed-loop filtering around a simulator, and physics-guided probabilistic denoising. The literature does not privilege one pattern universally; instead, each pattern addresses a different failure mode of monocular pose estimation.

6. Evaluation, recurring trade-offs, and limitations

Reported evaluations are heterogeneous, spanning Human3.6M, 3DOH, MPI-INF-3DHP, Physio2.2M, MoviCam, 3DPW, Fit3D, SportsPose, HumanEva-I, and MoYo. As a result, the literature reports task-specific gains rather than directly comparable rankings. In the Physio2.2M physiotherapy setting, the BlazePose-based PhysDynPose reduces BlazePose world 3D MPJPE from $G(\theta)$5 mm to $G(\theta)$6 mm with reset calibration and to $G(\theta)$7 mm with reused calibration, reduces average angle error from $G(\theta)$8 to $G(\theta)$9 or $q=[r;\theta]\in\mathbb{R}^{6+J}$0, and reduces bone-length variance from $q=[r;\theta]\in\mathbb{R}^{6+J}$1 mm$q=[r;\theta]\in\mathbb{R}^{6+J}$2 to $q=[r;\theta]\in\mathbb{R}^{6+J}$3 mm$q=[r;\theta]\in\mathbb{R}^{6+J}$4, with statistical tests reporting $q=[r;\theta]\in\mathbb{R}^{6+J}$5 improvements for most joints and angles (Leuthold et al., 7 Dec 2025). In the online physics-based setting of OSDCap on Human3.6M, the reported results are MPJPE $q=[r;\theta]\in\mathbb{R}^{6+J}$6 mm, MPJPE-G $q=[r;\theta]\in\mathbb{R}^{6+J}$7 mm, MPJPE-PA $q=[r;\theta]\in\mathbb{R}^{6+J}$8 mm, PCK $q=[r;\theta]\in\mathbb{R}^{6+J}$9, CPS $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$0 mm, GRP $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$1 mm, and Accel $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$2 mm/s$$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$3 (Le et al., 2024).

A common misconception is that adding physics uniformly improves every conventional reconstruction metric. The MoviCam results show a more specific trade-off. On non-flat sequences, 4DHumans reports MPJPE $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$4 and WA-MPJPE $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$5, whereas Ours (PhysDynPose) reports MPJPE $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$6 and WA-MPJPE $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$7; conversely, PhysDynPose improves W-MPJPE to $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$8 from $$M(q)\,\ddot q + h(q,\dot q)=\tau + J_c(q)^T\lambda,$$9 and RTE to $J_c(q)$0 from $J_c(q)$1, while also reducing foot sliding to $J_c(q)$2 mm from $J_c(q)$3 mm for 4DHumans (Aytekin et al., 23 Jul 2025). This indicates that scene-aware physics can preferentially improve world-frame consistency and physical plausibility even when local pose metrics do not uniformly decrease.

Another misconception is that physics necessarily appears only as a post-processing stage. The literature explicitly contests that view. Physics-guided reverse diffusion uses tracking discrepancies as denoising guidance (Ju et al., 2023); SMPL-Dynamics is designed as a differentiable inverse-dynamics layer that can be folded into learning or test-time refinement (Kim et al., 23 May 2026); and OSDCap makes physics part of the online state estimator itself (Le et al., 2024). At the same time, the older contact-and-dynamics pipeline remains a genuine post hoc optimizer, requiring 30–60 min for a 2 s clip and assuming single planar ground, no occlusions of feet, and no object, hand, or knee contact (Rempe et al., 2020). The field therefore contains both integrated and post-processing interpretations.

The reported limitations are similarly specific. OSDCap notes that the proxy character is approximate, external forces are only modeled at the feet, no explicit body-shape estimation is included, and ground-truth forces are unavailable, so force validity is assessed indirectly through pose (Le et al., 2024). The moving-camera PhysDynPose depends on SLAM, scene geometry, and a height map, with scale resolved by aligning the first two frames to ground-truth camera motion (Aytekin et al., 23 Jul 2025). The anatomy-driven BlazePose variant relies on per-subject bone-ratio calibration and on confidence-gated ratio updates (Leuthold et al., 7 Dec 2025). These are not incidental implementation details; they define the operational envelope of each formulation.

In aggregate, PhysDynPose names a research program whose central claim is narrower than generic “physical plausibility.” It is the claim that monocular human pose estimation becomes more useful when kinematic predictions are treated as inputs to mechanics-aware inference rather than as final outputs. The concrete realization of that claim varies—from SMPL inverse dynamics and Butterworth filtering, to contact-constrained QPs, to neural Kalman filters, to diffusion guidance, to anatomical ratio estimation—but the underlying objective remains the same: pose estimates that are not only geometrically close to image evidence, but also dynamically and biomechanically coherent.