Visual realism does not imply physical fidelity. GAUGE evaluates whether physics engines and video world models reproduce real-world physical behavior — and diagnoses which physical mechanisms they fail to reproduce.

GAUGE is a real-world-grounded benchmark for evaluating the physical fidelity of both numerical physics engines and generative video world models.

Instead of relying primarily on visual similarity or human judgments, GAUGE compares simulated or generated dynamics against controlled real-world experiments, using measured trajectories, calibrated physical parameters, uncertainty estimates, and task-specific physical observables.

GAUGE provides:

• 22 controlled physical task families
• ~1,560 real-world motion-capture trials
• Rigid, cable, textile, and volumetric deformable objects
• Experimentally calibrated physical parameters
• Millimeter-level trajectory measurements
• Physics-engine and video-world-model evaluation tracks
• Task-specific diagnostic metrics and standardized evaluation protocols

---

🔥 News

• 2026-09 — GAUGE code and evaluation toolkit to be released.
• 2026-08 — Real-world benchmark data released.
• 2026-08 — Physics-engine and video-world-model baselines released.

---

🌍 What is GAUGE?

Physics simulators are increasingly used for robot learning, policy evaluation, and large-scale synthetic data generation. Meanwhile, generative video models are emerging as implicit simulators capable of predicting future interactions.

But a fundamental question remains:

How physically faithful are these simulators to the real world?

A rollout can look realistic while still having the wrong friction, acceleration, momentum transfer, oscillation period, deformation, or energy behavior.

GAUGE addresses this problem by grounding evaluation in controlled real-world physical experiments.

Each task provides:

Real Experiment → Physical Measurement → Matched Simulation / Video Generation → Observation Extraction → Physics Evaluation

This allows GAUGE to go beyond asking whether a rollout looks plausible and instead ask:

1. Does the trajectory match real-world dynamics?
2. Does the motion obey the expected physical law?
3. Are the underlying physical parameters correct?
4. Is the behavior temporally stable?
5. Which physical mechanism causes the failure?

---

✨ What Makes GAUGE Different?

📏 Measurement-Grounded

GAUGE uses repeated real-world experiments rather than simulator-generated reference trajectories.

Experiments are recorded using a high-precision motion-capture system and paired with calibrated physical metadata.

🧩 Cross-Regime Physics

One benchmark covers multiple physical regimes:

Regime	Tasks	Representative Physics
Rigid Body	8	Collision, friction, momentum transfer, oscillation
Flexible Cable	1	Self-contact, stretching
Textile	6	Stretching, bending, friction, rapid dynamic
Volumetric Deformable Body	7	Stretching, compression, shear, twist, bending

⚙️ Physics Engines + 🌐 World Models

GAUGE evaluates two complementary classes of simulators under a common real-world foundation:

Simulation-Engine Track

Real experiment → matched simulated scene → simulated trajectory → sim-to-real evaluation

Video World Model Track

Initial image + prompt → generated video → motion extraction → physical-law evaluation

🔬 Diagnostic, Not Just Ranking

GAUGE is designed to identify why a simulator fails.

Evaluation includes trajectory discrepancy, temporal alignment, momentum transfer, oscillation, force-motion consistency, equation-form consistency, and physical-parameter accuracy.

---

🧪 Benchmark Tasks

GAUGE contains 22 controlled task families spanning rigid bodies, flexible cables, textiles, and volumetric deformable objects.

Category	Task	Video	Description	Evaluation	Subtasks	Material
Rigid Body	Slope Contact		Fall of a trirectangular tetrahedron	Slope collisions	1	Wood, Plastic
Rigid Body	Nonsmooth Contact		Fall of wedges and pyramids	Codimensional collisions	3	Wood, Plastic
Rigid Body	Slope Slider		Slide of a cube on a slope	Static and kinetic friction	3	Wood, Plastic, Metal
Rigid Body	Turntable		Slide of a cube on a turntable	Static and kinetic friction in a non-inertial frame	3	Wood, Plastic, Metal
Rigid Body	Bouncing Ball		Fall and bouncing of a ball	Rapid impact	1	Rubber
Rigid Body	Newton's Cradle		Collision of closely fitted balls	Momentum transfer	1	Metal
Rigid Body	Wall Breaking		Collision of a wrecking ball with a wall	Large-scale dense collisions	1	Wood
Rigid Body	Pendulum		Oscillation of a ball	Periodic motion	1	Metal
Flexible Cable	Rope Winding		Winding of a rope	Self-collision and stretch modulus	1	Rubber
Textile	Textile Stretching		Tensile deformation of a textile	Stretch modulus	1	R, S, U, O, L, N
Textile	Textile Bending		Natural sagging of a textile	Bending modulus	1	R, S, U, O, L, N
Textile	Textile Flinging		Fluttering of a textile	Motion under locally high acceleration	1	R, S, U, O, L, N
Textile	Funnel		Passage of textiles through a hole	Collision, friction	1	R, S, U, O, L, N
Textile	Rotating Ball		Motion of a textile driven by a rotating ball	Static and kinetic friction	1	R, S, U
Textile	Tablecloth Pulling		Pulling a textile from beneath rigid objects	Static and kinetic friction	1	Wood, Plastic, Metal
Volumetric Deformable Body	Foam Stretching		Stretching of an elastic cuboid	Stretching modulus	1	Soft, Hard
Volumetric Deformable Body	Foam Compression		Compression of an elastic cuboid	Compression modulus	1	Soft, Hard
Volumetric Deformable Body	Foam Shearing		Shearing of an elastic cuboid	Shear modulus	1	Soft, Hard
Volumetric Deformable Body	Foam Twisting		Twisting of an elastic cuboid	Twisting modulus	1	Soft, Hard
Volumetric Deformable Body	Foam Bending		Bending of an elastic cuboid	Bending modulus	1	Soft, Hard
Volumetric Deformable Body	Stick-Stack		Sliding of an elastic rod on a plane	Static and kinetic friction	1	Soft, Hard
Volumetric Deformable Body	Cantilever Beam		Free hanging of an elastic cantilever beam	Coupling with a large stiffness ratio	1	Soft, Hard

Textile material abbreviations: R = Rayon, S = Satin, U = Uniform cloth, O = Oxford fabric, L = Synthetic leather, N = Nylon taslan.

---

📦 Dataset

Each GAUGE task is grounded in repeated physical experiments.

The dataset contains approximately 1,560 motion-capture trials, together with calibrated physical properties and uncertainty estimates.

Real-World Measurements

For rigid objects:

local frame
        ↓
6-DoF pose trajectory over time

For textiles and volumetric deformable objects:

tracked surface markers
        ↓
position trajectory over time

Physical Metadata

Depending on the task and material, GAUGE provides experimentally characterized quantities including:

• geometry
• mass
• density
• friction
• restitution
• tensile stiffness
• shear stiffness
• bending stiffness
• Young's modulus
• Poisson's ratio

Project layout

GAUGE is split across a small constellation of repositories. This repo is the front door:

Component	Where it lives	What it provides
GAUGE (this repo)	InternRobotics/GAUGE	Reference baselines, scripts, project entry point
GAUGE-Dataset	🤗 GAUGE-Dataset	Scenes, task assets, physical metadata and measured trajectories (json format)

GAUGE/
├── scripts/                    # Evaluation scripts for GAUGE benchmark tracks
│   ├── simulation/             # Physics-engine simulation 
│   └── metrics/                # Trajectory and task-specific physical metrics and sim-to-real evaluation
│
├── media/                      # README figures and task visualizations
├── LICENSE                     # Project license
└── README.md                   # Benchmark overview, setup, evaluation, and usage

---

⚙️ Simulation-Engine Track

The simulation-engine track evaluates whether a numerical simulator reproduces measured real-world dynamics under matched initial conditions and physical parameters.

GAUGE currently evaluates representative tasks with:

Physics Engine	Rigid	Textile	Deformable
IsaacSim	PhysX	Surface FEM	FEM
Genesis	Default	PBD	Explicit MPM
Newton	Mujoco	VBD	Implicit MPM

Physical parameters are initialized from GAUGE's experimental calibration rather than task-specific simulator fitting.

Generalized Trajectory

Different object classes are represented using task-appropriate observables:

Object	Representation
Rigid body	Position
Textile	Marker-wise Gaussian curvature
Volumetric deformable body	Local mesh-face area

This allows different physical regimes to be compared under a common trajectory-evaluation framework while retaining physically meaningful state representations.

Metrics

Core metrics include:

• RMSE — Frame-aligned discrepancy between simulated and mean real-world trajectories.

• DTW — Dynamic Time Warping discrepancy, allowing monotonic temporal alignment.

Task-specific metrics additionally include:

• LSD — Longest Stationary Duration
• MTE — Momentum Transfer Efficiency
• PD — Period Duration
• EL — Energy Loss

---

🌐 Video World Model Track

The world-model track asks a different question:

Has a generative model learned the underlying physics, or only how physical motion should look?

Generated objects are segmented and tracked frame by frame, after which pixel trajectories are converted into physical coordinates.

World Model	Cosmos3-Nano	Cosmos3-Super	Wan-2.2	Wan-2.7	Seedance-2.0	Genie3
Positive Prompt	✓	✓	✓	✓	✓	✓
Negative Prompt	✓	✓	✓	✓	×	×

Metrics

GAUGE separates physical-law structure from physical-parameter accuracy.

Physical Law Structure

• Dynamic Error (DE) — Measures consistency between predicted motion and expected force–motion relationships.

• Coefficient of Determination (R²) — Measures how well the expected governing equation explains the generated trajectory.

• Quadratic Form Improvement (QFI) — Tests whether unexpected higher-order curvature is needed to explain motion that should follow a simpler physical relationship.

Physical Parameters

Task-specific parameters such as:

• acceleration
• momentum-transfer efficiency
• oscillation period
• damping

are recovered from generated trajectories and compared with real measurements.

This distinction is important because a generated trajectory may fit the form of a physical law while still recovering the wrong physical parameters.

---

📊 Benchmark Results

Physics Engines

We benchmark representative GAUGE tasks using:

Engine	Rigid	Textile	Volumetric
Isaac Sim	✓	✓	✓
Genesis	✓	✓	✓
Newton	✓	✓	✓

A central finding is that no single physics engine is uniformly faithful across all physical regimes.

Standard contact and sliding behaviors can often be reproduced reasonably well, while substantially larger discrepancies appear in:

• impulsive contact,
• high-acceleration textile motion,
• volumetric deformation.

Video World Models

GAUGE additionally evaluates representative image-to-video and interactive world models on rigid-body physical tasks.

The results reveal an important distinction:

A model can generate motion with the correct-looking equation form while predicting incorrect physical parameters.

Examples include incorrect acceleration, momentum transfer, oscillation period, and damping despite visually plausible motion.

---

🔎 Understanding Your Results

GAUGE is intended as a diagnostic benchmark, not only a leaderboard.

A useful interpretation is:

High visual quality
      │
      ├── Correct equation form?
      │       │
      │       ├── No  → structural physics failure
      │       │
      │       └── Yes
      │            │
      │            ├── Correct parameters?
      │            │       └── No → quantitative physics failure
      │            │
      │            └── Correct temporal behavior?
      │                    └── No → dynamics / stability failure
      │
      └── Compare against real experimental uncertainty

The goal is therefore not simply to answer:

Which simulator scores highest?

but rather:

Which physical mechanisms does each simulator reproduce faithfully, and where does it deviate from reality?

---

🛣️ Roadmap

Future versions of GAUGE aim to extend the benchmark toward:

• more materials and broader physical-parameter ranges,
• additional rigid/deformable interaction regimes,
• fluid dynamics,
• fluid–rigid interaction,
• fluid–soft-body interaction,
• 3D world-model evaluation,
• reconstructed point-cloud and mesh metrics,
• strain, curvature, bending-energy and energy-dissipation analysis.

---

📝 Citation

If you find GAUGE useful in your research, please consider citing:

@article{Wang2026,
  title={GAUGE: A Measurement-Grounded Benchmark for Physical Fidelity in Simulation Engines and Video World Models},
  author={Wang, Shuai and Feng, Yaxin and Jiang, Xuekun and Tian, Shihan and others},
  year={2026},
  copyright = {arXiv.org perpetual, non-exclusive license},
  doi       = {10.48550/ARXIV.2608.05948},
  keywords  = {Artificial Intelligence (cs.AI), Computer Vision and Pattern Recognition (cs.CV), Robotics (cs.RO), FOS: Computer and information sciences},
  publisher = {arXiv},
}

Replace with the final BibTeX entry from the released paper.

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.