Physical Stance in Multidisciplinary Research

• Physical stance is a multifaceted term that encompasses physically plausible postures and dynamics across vision, biomechanics, robotics, and philosophy.
• It integrates stringent constraints such as contact consistency, stability, and accurate force modeling to enhance real-world application fidelity.
• Evaluation relies on measurable criteria like joint errors, center-of-mass deviation, and balance metrics to assess dynamic physical plausibility.

Physical stance is a polysemous research term rather than a single standardized construct. In computer vision and motion capture, it denotes pose and trajectory estimates that remain compatible with dynamics, gravity, friction, contact, scene geometry, and world-coordinate consistency. In biomechanics, it denotes the control of upright posture, body sway, and trunk stabilization under perturbation. In robotics, it denotes the stance or contact regime through which support forces, compliance, and angular-momentum regulation are realized. In philosophy of mind, “Physical Stance” denotes the interpretation using the language of physics, contrasted with a “Causal Stance.” This suggests that the common thread is not a particular posture, but physically grounded description at the level relevant to a domain (Aytekin et al., 23 Jul 2025, Lippi et al., 2022, Fahmi et al., 2019, Ohmura et al., 6 Apr 2026).

1. Domain-specific meanings

Across recent arXiv literature, “physical stance” appears in several technically distinct but structurally related senses.

Context	Operational meaning	Representative papers
Human pose estimation	Physically plausible 3D pose and motion respecting contacts, scene geometry, and world coordinates	(Aytekin et al., 23 Jul 2025, Shimada et al., 2021, Tripathi et al., 2023, Leuthold et al., 7 Dec 2025)
Biomechanics	Upright stance, body sway, and trunk stabilization under perturbation	(Lippi et al., 2022, Drama et al., 2020)
Legged robotics	Contact-consistent support, compliance adaptation, and pendular force organization	(Fahmi et al., 2019, Dhawale et al., 2018, Lyu et al., 7 May 2026)
Ergonomics and XR	Literal user posture such as sitting, standing, or kneeling	(Kojić et al., 26 Feb 2026, Białecki et al., 2024, Sreenivasan et al., 2024)
Philosophy of mind	Interpretation using the language of physics	(Ohmura et al., 6 Apr 2026)

Recurring state variables and observables include contact forces, center of mass (CoM), center of pressure (CoP), support geometry, friction constraints, and stability over time. A frequent methodological contrast is between kinematic accuracy and physical plausibility: several papers explicitly show that a low joint-position error does not by itself ensure balance, contact consistency, or stable motion.

2. Physics-based human pose estimation

In monocular human pose estimation, physical stance is defined most explicitly as a physically plausible trajectory rather than a purely kinematic sequence. The central motivation is that kinematic estimators can produce ground penetration, floating, sliding, drift, and incoherent global motion when the camera moves or the environment is non-planar (Aytekin et al., 23 Jul 2025).

“Physics-based Human Pose Estimation from a Single Moving RGB Camera” introduces PhysDynPose as a two-stage plug-and-play pipeline. A state-of-the-art kinematics estimator, specifically 4DHumans, predicts body pose and shape in a root-relative frame, while DROID-SLAM estimates the dynamic camera trajectory. The human pose is then lifted into world coordinates and refined by a scene-aware physics optimizer in PyBullet. Two stated enhancements are scene geometry integration through a height map and explicit root supervision to reduce drift. Foot-ground penetration is tested against the terrain height,

$$r_{\text{foot}, y} < h(r_{\text{foot}, x}, r_{\text{foot}, z}),$$

so non-flat geometry replaces the flat-ground assumption used in many earlier pipelines (Aytekin et al., 23 Jul 2025).

A related line of work embeds physical awareness directly into differentiable motion capture. “Neural Monocular 3D Human Motion Capture with Physical Awareness” presents a fully differentiable system combining a neural proportional-derivative controller, an explicit rigid body dynamics model, and an optimization layer that enforces no foot-floor penetration as a hard constraint. Its purpose is not only smoother motion, but physically principled reconstruction under challenging in-the-wild conditions and camera variation (Shimada et al., 2021).

Other formulations replace full simulation with differentiable biomechanical surrogates. “3D Human Pose Estimation via Intuitive Physics” develops IPMAN from a pressure heatmap, the CoP inferred from that heatmap, and the SMPL body’s CoM. The CoP is defined as

$$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$

and stability is encouraged by a differentiable loss penalizing the distance between the projected CoM and the support region. The paper reports poses that are 14.8% more stable than baseline methods in post-hoc Bullet simulations, while improving accuracy for static poses and not hurting dynamic ones (Tripathi et al., 2023).

A more anatomy-centered usage appears in “Physics Informed Human Posture Estimation Based on 3D Landmarks from Monocular RGB-Videos.” There, physical stance means anatomically consistent 3D posture recovered by fusing BlazePose 3D and 2D outputs with weighted optimization, bone-length penalties, biomechanical models, and a Kalman filter for subject-specific bone ratios. On Physio2.2M, the method reports a 10.2 percent reduction in 3D MPJPE, a 16.6 percent decrease in errors of angles between body segments, and a reduction in bone length variance from 370 mm² to 21 mm² (Leuthold et al., 7 Dec 2025).

3. Benchmarks and evaluation criteria

The evaluation of physical stance has shifted from ad hoc visual judgment toward datasets and metrics that target global consistency, contact, and stability. MoviCam is introduced as the first non-synthetic dataset containing ground-truth camera trajectories of a dynamically moving monocular RGB camera, scene geometry, and 3D human motion with human-scene contact labels. Its stated role is to expose failure modes that remain hidden in static-camera, synthetic, or ground-truth-free evaluations (Aytekin et al., 23 Jul 2025).

A complementary development is simulation-based plausibility measurement. “Measuring Physical Plausibility of 3D Human Poses Using Physics Simulation” introduces CoM Distance (CD) and Pose Stability Duration (PSD$_T$):

$$\text{CD} = \frac{1}{T} \sum_{t=1}^{T} \|\mathbf{c}^k_t - \mathbf{c}_t\|_2,$$

$\text{PSD}_T = \min(T - N, t_F).$

CD measures deviation between the simulated CoM trajectory and the kinematic reference; PSD$_T$ measures how long a simulated body remains stable before catastrophic failure or out-of-support behavior. The paper explicitly argues that low per-frame error does not guarantee plausible or stable motion (Louis et al., 6 Feb 2025).

The numerical comparison in that study makes the point sharply. PoseFormer has MPJPE 42.5, FS% 4.4, GP 0.30, CD 36.2, and PSD$_{100}$ 64.8, whereas the multi-view baseline has MPJPE 55.6, FS% 2.6, GP 0.26, CD 27.3, and PSD$_{100}$ 70.6. NeuralPhysCap has MPJPE 81.6, FS% 29.7, GP 2.62, CD 29.6, and PSD$_{100}$ 62.3. These values formalize a recurring theme in the literature: kinematic fidelity and physical plausibility are correlated only partially, and sometimes weakly (Louis et al., 6 Feb 2025).

4. Human postural control and running

In experimental biomechanics, physical stance is studied as a control problem for upright balance. “Human body-sway steady-state responses to small amplitude tilts and translations of the support surface” treats the response as a frequency-response problem. Under small perturbations, body sway and ankle angle had correlation $r = 0.9894$, so a single inverted pendulum (SIP) was used. Frequency response functions were computed as

$$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$0

The paper reports significant differences between eyes-open and eyes-closed conditions across all tested conditions, while superposition of tilt and translation did not significantly alter FRFs in most cases; only 5 of 24 tested comparisons were significant. Simulations further suggest that modulation of ankle stiffness explains the observed interaction for unbalanced stimuli (Lippi et al., 2022).

In running, the relevant stance problem is not quiet balance but trunk stabilization under dynamic ground reaction forces. “Postural Stability in Human Running with Step-down Perturbations” analyzes the virtual point (VP) mechanism and finds that the VP is consistently below the CoM, i.e. a VPB, with mean vertical location approximately $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$1 cm. The trunk moves forward during single stance, consistent with VPB dynamics. Visible and camouflaged step-down perturbations alter VP location and reduce the fit quality $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$2, but VP remains below the CoM in all running conditions. The matched TSLIP simulations suggest that the VPB mechanism can encounter step-down perturbations and return the system to its initial equilibrium state (Drama et al., 2020).

These studies distinguish two common misconceptions. First, stance is not only static posture; it also includes dynamically stabilized locomotor phases. Second, combined perturbations do not automatically imply strongly nonlinear response coupling; in the small-amplitude regime studied for upright stance, disturbance-specific processing dominates.

5. Robotic stance, terrain interaction, and pendular structure

In legged robotics, stance is the contact configuration through which support forces are generated and regulated. “STANCE: Locomotion Adaptation over Soft Terrain” integrates whole-body control with online terrain compliance estimation. Contact is modeled with a Kelvin-Voigt law, terrain stiffness and damping are estimated per foot using weighted least-squares regression, and the estimated parameters are fed back into the controller. The paper reports online adaptation to rigid and soft terrain without pre-tuning, successful transitions between terrains, per-foot differentiation of compliance, and computational cost of less than 2 ms per control cycle at 250 Hz (Fahmi et al., 2019).

A more reduced mechanical model appears in “How to run on rough terrains.” There, the runner is a rigid body in the sagittal plane with passive impulsive collision and active impulsive push-off. Open-loop strategies suffice on step-like terrains with piecewise flat surfaces that randomly vary in height, but orientational stability is lost on rough terrains whose slope and height vary randomly unless tangential collisions are avoided. Leg retraction that matches foot and ground speed at touchdown is the main mechanism for recovering orientational stability. The paper derives the dimensionless stability criterion

$$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$3

linking robustness to allowable tilt and perturbation-induced orientation error (Dhawale et al., 2018).

At the centroidal-control level, “On the Emergence of Pendular Structure in Multi-Contact Locomotion” makes the link between stance geometry and LIPM-like behavior explicit. With full-rank stance, the optimum of a centroidal OCP penalizing the rate of angular momentum drifts toward a pendular force pattern at a rate set by the singular values of the moment Jacobian. With $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$4 stance, as in trot, the friction cone imposes a lower bound on $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$5 that weight tuning cannot remove, and a non-smooth feasibility kink appears at a critical horizontal acceleration written in closed form. On the pendular manifold, classical ZMP and DCM relations are recovered. Experiments on a point-mass quadruped and Unitree Go1 show that the geometric constant predicted from the SVD matches experiment to within 16% in the full-rank case (Lyu et al., 7 May 2026).

A related occupational-robotics use of stance appears in exoskeleton evaluation. “Exoskeleton-Assisted Balance and Task Evaluation During Quiet Stance and Kneeling in Construction” models quiet stance as a triple-link inverted pendulum and kneeling as a double-link inverted pendulum, both with LQR control. The paper reports that center of pressure sway area was reduced up to 62% in quiet stance and 39% in kneeling when knee exoskeleton assistance was provided in high-elevation VR/MR worksites, and argues that the knee joint is crucial for balance control in both postures (Sreenivasan et al., 2024).

6. Posture, ergonomics, and user experience

Some recent work uses physical stance in the literal sense of adopted posture. In XR climbing, stance means sitting or standing during exposure to virtual heights. A user study with 25 participants found that seated participants experienced slightly higher immersion but were also more susceptible to motion sickness compared to those standing. For IPQ realism, the reported means were 2.24 for sitting and 1.87 for standing, with $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$6 and $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$7. The highest nausea occurred in the City Nighttime environment with mean 43.3 and standard deviation 31.1, whereas Gym 1 had mean 13.4 and standard deviation 16.3 (Kojić et al., 26 Feb 2026).

In esports biomechanics, stance is operationalized through stabilometry. “Esports Training in StarCraft II: Stance Stability and Grip Strength” compares 11 esports players with 24 physical-education students. The main balance outcome is CoP path length. There is no significant group difference in bipedal stance, with $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$8, $$\text{CoP} = \frac{\sum_i \rho_i \mathbf{v}_i}{\sum_i \rho_i},$$9, and $_T$0, but esports players are significantly worse in single-leg stance, with $_T$1, $_T$2, and $_T$3. Grip strength is also lower in the esports group across repeated attempts (Białecki et al., 2024).

These uses are narrower than those in physics-based pose estimation or robotics, but they preserve the same structural concern: stance is the posture through which support, control, and instability become measurable.

7. Philosophical usage and methodological caution

A distinct usage appears in philosophy of mind. “Causal Stance” defines the Physical Stance as the interpretation using the language of physics, and the Causal Stance as the interpretation using the language of causation. Within this framework, physical determinism belongs to the Physical Stance, whereas physical causal closure is defined only within the Causal Stance. The central claim is that physics describes lawful state evolution, but does not itself define asymmetric, intervention-based causation; accordingly, physical determinism and physical causal closure should not be equated (Ohmura et al., 6 Apr 2026).

A broader methodological caution is developed in “Frenemies with Physicality: Manufacturing Manifold Metrics.” That paper argues that physicality is multifaceted and context-dependent, and that criteria such as energy conditions may be appropriate in one setting and misplaced in another, especially in analogue space-times and toy models. The warning generalizes beyond relativity: importing a notion of “physicality” or “physical stance” outside its native theory can distort rather than clarify analysis (Schuster, 2023).

A plausible implication is that the most productive use of physical stance is always theory-relative. In pose estimation it means dynamics- and scene-consistent reconstruction; in biomechanics, balance and disturbance rejection; in robotics, feasible contact force generation under geometric and frictional constraints; in posture-oriented HCI, the bodily configuration that shapes perception and discomfort; and in philosophy, a descriptive stance defined by the language of physics rather than the language of causation.