Latent Dynamics · 2026-10-20
Beyond Geometric Closure: Measuring and Diagnosing Mechanical Consistency in Visual Dynamics
Qibin Gu, Yunfeng Yan
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors propose a six-criterion framework to measure mechanical consistency in visual dynamics, auditing geometric admissibility, reaction-free actuator effort, and work–energy balance in SI units. They evaluate KeyCLD and CON on 343 planar assemblies and V-JEPA+AC and LingBot-VA on 100 spatial assemblies, showing that lower image MSE does not imply native mechanical consistency, with calibrated replay revealing forecast errors far above measurement error.
Why it matters
Visual prediction can look accurate while producing motion that violates mechanical constraints. By decomposing errors into geometric admissibility, actuator effort, and work–energy balance (and using reference replay plus interventions), the authors aim to diagnose where failures come from: readout/localization, numerical differentiation/integration, or learned force/constitutive response. This links visual dynamics to physically meaningful actuator demands and provides continuous SI tolerance scoring.
Method
- Introduce independent SI audit criteria for rigid closed chains: closure/admissibility, velocity and curvature-consistent acceleration, torque (reaction-free actuator effort), energy/work–energy residual, and point tracking error vs reference trajectory.
- Use reference replay with exact-state and measurement-path controls to isolate numerical error and separate coordinate readout, integration, and learned force/model errors.
- Apply targeted interventions (resolution matching, exact-state controls, RK4 vs Euler, and CAD/IK reconstruction) to determine which components improve visual prediction vs mechanical consistency.
Limitation
The paper text states that despite improvements (e.g., image MSE and some geometric components), neither planar model attains valid native geometry, and spatial interventions restore geometric admissibility while actuator residuals remain large (coverage for strict joint qualification remains zero for both spatial and planar native outputs).
Abstract (from arXiv)
Visual accuracy leaves open whether predicted motion obeys dynamics. We introduce a framework for measuring and diagnosing mechanical consistency in visual dynamics. An independent SI audit of rigid closed chains separates geometric admissibility, reaction-free actuator effort, and work--energy balance. Continuous scoring measures progress against fixed engineering targets, while reference replay quantifies measurement error. With three seeds per setting, we evaluate KeyCLD and CON across 343 planar assemblies at 64/256 pixels and V-JEPA + AC and LingBot-VA on 100 shared spatial test assemblies. CON reduces image MSE by 37.6%/37.2% relative to persistence, yet neither planar model attains valid native geometry. At 256 pixels, calibrated RGB replay places forecast errors far above measurement error. With exact states, KeyCLD preserves geometry and RK4 training raises its score from 52.65 to 64.30. In both spatial pathways, CAD/IK reconstruction restores geometry while mean actuator residuals exceed five times the engineering targets. The interventions distinguish gains in visual prediction and geometric admissibility from remaining force and energy errors.
Related papers
- World Models' Last Exam in Physics
- PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements
- PhysicsLENS: Diagnosing Physical Property Blindness in Video Generation Models
- MotionCore: A Latent Neural Simulator for Video Generation with Rigid-Body Dynamics
- Diagnosing and Recovering from Observation-Space Shift at Long-Horizon Skill Seams
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-17 · Indexed by Pub-AI: 2026-10-20 UTC
Beyond Geometric Closure: Measuring and Diagnosing Mechanical Consistency in Visual Dynamics
Preprint posted: 2026-10-17 · Original record ↗
- arxiv
- 2610.24619
Explicit source versions (1)
- v3 · 2026-10-21
- Preprint available
Source date: 2026-10-17 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
- Explicit source version detected · v3
Source date: 2026-10-21 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
Abstract · arxiv
Visual accuracy, geometric validity, and agreement with applied forces are distinct requirements for visual dynamics. We introduce a diagnostic framework that connects reaction-free SI measurements to controlled interventions on coordinate readout, training integration, and actuator response. Reference replay calibrates the measurement path, while continuous scoring exposes progress against fixed engineering targets. With three seeds per setting, we evaluate KeyCLD and CON across 343 planar assemblies at 64/256 pixels and V-JEPA + AC and LingBot-VA on 100 shared spatial test assemblies. CON reduces image MSE by 37.6%/37.2% relative to persistence, yet neither planar model attains valid native geometry. At 256 pixels, calibrated RGB replay places forecast errors far above the reference measurement floor. Exact-state KeyCLD has identical geometric pass rates under Euler and RK4 training, yet its score rises from 52.65 to 64.30 as force and energy errors decrease. In both spatial pathways, CAD/IK reconstruction restores geometry while mean actuator residuals exceed five times the engineering targets. These diagnoses distinguish coordinate correction, numerical improvement, and persistent force-response errors, guiding targeted improvements to visual dynamics.
