PhyGround: Framework for Physical Evaluation

• PhyGround is a dual-use concept defined as either a benchmark assessing law-specific physical plausibility in video generation or as a physics-informed method for human motion analysis.
• The benchmark decomposes physical evaluation into 13 distinct laws across solid-body, fluid, and optical domains, using both controlled human studies and an automated judge (PhyJudge-9B).
• The human-motion application of PhyGround estimates ground reaction forces from motion capture data by applying Newton–Euler constraints and PD controllers for improved trajectory reconstruction.

Searching arXiv for the exact term and closely related uses to ground the article in current papers. PhyGround most commonly denotes a criteria-grounded benchmark for evaluating physical reasoning in generative world models, particularly text+image-to-video systems, through explicit physical-law diagnostics rather than generic video-quality metrics (Lin et al., 11 May 2026). In the current literature, the same name also appears in a second, distinct sense as shorthand for “Physics-informed Ground Reaction Dynamics from Human Motion Capture,” a method for estimating human ground reaction forces from motion data under Newton–Euler constraints (Le et al., 2 Jul 2025). The dominant benchmark usage concerns whether generated videos obey observable laws of solid-body mechanics, fluid dynamics, and optics; the ground-reaction usage concerns physically informed inference of forces and trajectories from human kinematics.

1. Scope and nomenclature

In its benchmark sense, PhyGround was introduced to address a specific evaluation problem in generative world models: high visual fidelity does not guarantee adherence to physical rules. The benchmark formulation identifies three limitations in prior physics-focused video evaluation: lack of law-specific diagnosis, uncontrolled human evaluation biases, and non-reproducible, general-purpose auto-evaluation (Lin et al., 11 May 2026). These limitations motivate a design in which each physical law is scored separately, human judgments are collected under controlled conditions, and automated evaluation is performed by an open physics-specialized judge rather than an opaque general-purpose API.

A common misconception is to treat PhyGround as a video generator or a learned simulator. In the benchmark paper, it is instead an evaluation stack composed of a prompt suite, a law taxonomy, a human-study protocol, and an automated judge (Lin et al., 11 May 2026). Its purpose is diagnostic: to expose which laws a model violates and how severely, rather than to improve generation directly.

The name is not unique across subfields. In biomechanics and motion modeling, “PhyGround” refers to a physics-informed method for human ground reaction dynamics estimation from motion capture (Le et al., 2 Jul 2025). Related but differently named work also includes “visual grounding” of learned particle-based physical models from video, where grounding means reconciling visual observations with a learned dynamics prior (Li et al., 2020). This dual usage makes terminological disambiguation necessary in technical writing.

2. Benchmark construction and law taxonomy

PhyGround, as a benchmark, contains 250 curated prompts. Each prompt is augmented with an expected physical outcome, and the first frame is provided as an image condition under a unified text+image-to-video format. The prompts are drawn from VideoPhy-2, Physics-IQ, OpenVid, and WorldModelBench, and are filtered for unique physical outcome, visual verifiability in a short video clip, concise everyday scenarios under 50 words, domain alignment with fluid, rigid-body, or optical phenomena, and core physical drivers that are passive and observable rather than driven by human agency (Lin et al., 11 May 2026).

The benchmark decomposes physical reasoning into 13 laws across three domains. Each law is operationalized by a human rating question, and for the automated judge each law is further broken into 2–3 binary sub-questions. All evaluations use a 1–5 Likert scale, where 1 is fully implausible and 5 is fully plausible (Lin et al., 11 May 2026).

Domain	Law	Human rating question
Solid-Body Mechanics	Gravity	Do objects and liquids move in accordance with gravity (unsupported objects fall downward)?
Solid-Body Mechanics	Inertia	Do objects maintain state of motion (remain stationary or continue moving) absent external force?
Solid-Body Mechanics	Momentum	Are post-collision directions physically plausible (conservation of momentum, $F=ma$)?
Solid-Body Mechanics	Impenetrability	Do solid objects avoid passing through one another?
Solid-Body Mechanics	Collision	Do objects exhibit plausible impact response proportional to applied force?
Solid-Body Mechanics	Material	Do objects respond according to apparent material properties (rigid vs. deformable)?
Fluid Dynamics	Buoyancy	Do objects sink or float consistent with density differences (Archimedes’ principle)?
Fluid Dynamics	Displacement	Does liquid level rise plausibly when volume is added or objects submerged?
Fluid Dynamics	Flow dynamics	Does liquid flow, spread, and drain in a physically plausible pattern (e.g. continuity equation $\nabla \cdot v = 0$)?
Fluid Dynamics	Boundary interaction	Does liquid splash, deflect, or split plausibly upon hitting a boundary?
Fluid Dynamics	Continuity	Does liquid mass remain continuous (no spontaneous gaps or creations)?
Optical Physics	Reflection	Do reflective surfaces display content consistent with scene geometry (Snell’s law: $n_1 \sin \theta_1 = n_2 \sin \theta_2$)?
Optical Physics	Shadow	Are shadows consistent in direction, placement, and motion relative to light sources and objects?

This taxonomy is central to the benchmark’s design. Rather than assigning a single “physics” score to an entire video, PhyGround isolates distinct observable failure modes. A model may satisfy flow direction while violating continuity, or produce plausible collisions while violating impenetrability. The benchmark therefore supports per-law diagnostics rather than only aggregate rankings (Lin et al., 11 May 2026).

3. Human-study protocol and reliability

PhyGround evaluates eight modern video generation models through a large-scale, quality-controlled human study grounded in social-science experimental design. The study recruited 459 annotators, consisting of 35 physics-educated experts and 424 broadly trained non-experts, under an IRB-approved, voluntary extra-credit scheme (Lin et al., 11 May 2026). Before quality control, the study produced 5,796 complete annotations and 37,408 fine-grained labels; after quality control, 352 annotators remained, yielding 4,576 complete video evaluations and approximately 29,500 fine-grained labels.

The protocol includes several explicit controls. Viewing was desktop-only, with an average of 12.6 videos per annotator. Video assignment was unbiased, presentation order was randomized, model identities were hidden, and the task framing did not reveal expected results. Annotators also completed a training module with worked examples that separated visual plausibility from physical correctness (Lin et al., 11 May 2026). These controls are intended to reduce fatigue, ordering effects, carryover effects, and demand characteristics.

Quality control was conducted in two rounds using multiple signals: score consistency, per-video copy-paste rate across dimensions, peer agreement measured by mean absolute error with co-annotators, and behavioral engagement indicators such as page dwell time and video play count (Lin et al., 11 May 2026). The retained annotations exhibited strong ranking stability. Split-half model-ranking correlations were reported as Spearman’s $\rho = 0.93$ for General, $0.91$ for Physics, and $0.93$ for Overall; Kendall’s $\tau_b$ was approximately $0.83$ across views. Pairwise preference agreement for total score was $73.0\%$, compared with $70\%$ in WorldModelBench and $\nabla \cdot v = 0$0 in VideoPhy-2 (Lin et al., 11 May 2026).

These reliability statistics matter because PhyGround is explicitly framed as a benchmark for law-specific physical judgment, not merely preference testing. The high split-half correlations indicate that model rankings are stable under subsampling of the annotator pool, which supports the validity of subsequent automated-judge training.

4. PhyJudge-9B and reproducible automated evaluation

To support reproducible large-scale evaluation, PhyGround introduces PhyJudge-9B, a 9B-parameter physics-specialized VLM judge fine-tuned on the human annotations with LoRA adapters (Lin et al., 11 May 2026). The base model is an off-the-shelf 9B multimodal LLM with a frozen ViT vision encoder and a trainable vision–language aligner. The LoRA configuration uses rank 32, alpha 64, and dropout 0.05, applied to all linear layers, including queries, keys, values, outputs, and MLPs. Training uses AdamW with $\nabla \cdot v = 0$1, $\nabla \cdot v = 0$2, weight decay 0.1, learning rate $\nabla \cdot v = 0$3 for LoRA and $\nabla \cdot v = 0$4 for the aligner, a cosine schedule with 5% warmup, bfloat16, and DeepSpeed ZeRO-2. The data comprise approximately 30K samples of video frames at 4 fps, a single-criterion prompt with sub-questions, and an aggregated human score $\nabla \cdot v = 0$5 (Lin et al., 11 May 2026).

The training objective is token-level negative log-likelihood:

$\nabla \cdot v = 0$6

Judge bias is evaluated by the absolute grouped relative bias

$\nabla \cdot v = 0$7

computed per dimension or law and then macro-averaged. Overall judge bias is defined as

$\nabla \cdot v = 0$8

On held-out Veo-3.1 outputs, PhyJudge-9B achieved General bias $\nabla \cdot v = 0$9, Physics bias $n_1 \sin \theta_1 = n_2 \sin \theta_2$0, Overall bias $n_1 \sin \theta_1 = n_2 \sin \theta_2$1, and signed bias $n_1 \sin \theta_1 = n_2 \sin \theta_2$2. The corresponding numbers for Gemini-3.1-Pro were $n_1 \sin \theta_1 = n_2 \sin \theta_2$3, $n_1 \sin \theta_1 = n_2 \sin \theta_2$4, $n_1 \sin \theta_1 = n_2 \sin \theta_2$5, and $n_1 \sin \theta_1 = n_2 \sin \theta_2$6. Claude Opus 4.7 had Overall bias approximately $n_1 \sin \theta_1 = n_2 \sin \theta_2$7 under chain-of-thought prompting, and Qwen3.5-9B had approximately $n_1 \sin \theta_1 = n_2 \sin \theta_2$8 (Lin et al., 11 May 2026). The paper also reports that adding the per-law sub-questions to judge prompts reduces bias sharply; for example, Gemini’s fluid bias drops from approximately $n_1 \sin \theta_1 = n_2 \sin \theta_2$9 to $\rho = 0.93$0. Frame sampling rates from 1 to 8 fps alter judge bias only slightly, and 4 fps is adopted as the default for efficiency and accuracy.

The reproducibility claim rests on open release rather than solely on reported metrics. The project provides the 250 annotated prompts with first-frame images, raw and QC-filtered human annotations, per-law scores, the PhyJudge-9B checkpoint and LoRA weights, evaluation code, prompt templates, and sub-question definitions (Lin et al., 11 May 2026). This directly addresses the criticism that closed-source judges cannot be audited or version-pinned.

5. Other research usages of “PhyGround”

A distinct use of the same name appears in “Physics-informed Ground Reaction Dynamics from Human Motion Capture,” where PhyGround denotes a method for estimating human ground reaction forces directly from motion capture using physics laws and computational simulation as constraints (Le et al., 2 Jul 2025). The method models the body as a single rigid mass $\rho = 0.93$1 acted on by ground reaction force $\rho = 0.93$2 and gravity $\rho = 0.93$3, with

$\rho = 0.93$4

Ground reaction force is approximated through a Proportional–Derivative controller,

$\rho = 0.93$5

and trajectory simulation is performed with a forward Euler scheme. The workflow extracts root translation from SMPL pose parameters, computes physics-based GRF candidates, selects PD gains by minimizing trajectory reconstruction error, and trains a network with 4 one-dimensional temporal convolution layers and 3 fully connected layers under a combined force-plate and physics loss. On GroundLink, the method reports average vertical GRF MSE of 0.09 | 0.06 versus 0.11 | 0.06 for GroundLink, and average vertical root position error of 14.69 ± 13.37 versus 38.43 ± 22.56, corresponding to an 18% reduction in left-foot vGRF error and a 62% reduction in trajectory error (Le et al., 2 Jul 2025).

This motion-centric PhyGround is closely related to GroundLink, a dataset of synchronized kinematics, ground reaction force, and center of pressure measurements. GroundLink contains 368 processed motion trials, approximately 1.59 million recorded frames, and 19 different movements, and GroundLinkNet serves as a benchmark neural model for predicting GRFs and CoPs from motion (Han et al., 2023). In the GroundLink discussion, one proposed extension is integration into physics-based simulators for unseen environments, explicitly labeled “PhyGround,” which indicates that the term also functions as a broader label for physically grounded contact modeling in graphics, biomechanics, and robotics.

A third, conceptually adjacent line of work is “Visual Grounding of Learned Physical Models,” which is not named PhyGround but addresses a similar problem of coupling perception to learned dynamics (Li et al., 2020). Li et al. describe a model with a visual prior $\rho = 0.93$6, an inference module $\rho = 0.93$7, and a dynamics prior $\rho = 0.93$8 based on DPI-Net. The system predicts particle-based representations from RGB observations, refines positions, rigidness, and global physical parameters, and then performs one-step-ahead dynamics prediction. It is evaluated on rigid objects, deformable materials, and fluids, and is reported to infer physical properties within a few observations, enabling adaptation to unseen scenarios (Li et al., 2020). In this setting, “grounding” refers to constraining learned physical variables by visual evidence.

6. Significance, misconceptions, and limitations

The principal significance of PhyGround in generative-model evaluation lies in its shift from domain-level or binary judgments to criteria-grounded, law-specific scoring. This makes benchmark outputs diagnostically useful: failures can be localized to gravity, continuity, reflection, or other observable laws rather than collapsed into a single undifferentiated “physics” score (Lin et al., 11 May 2026). A plausible implication is that this design is more suitable for model debugging and evaluator auditing than prior benchmarks centered on coarse domain labels.

Another misconception is that generic perceptual metrics are adequate proxies for physical plausibility. The PhyGround benchmark rejects that assumption explicitly: FVD, SSIM, and CLIP-similarity measure fidelity rather than physical correctness, while general-purpose VLM judges are not specialized for per-law diagnosis and may not be reproducible (Lin et al., 11 May 2026). The benchmark’s use of checklist-style sub-questions and an open judge is intended to make evaluation auditable at the level of individual physical laws.

Name ambiguity is itself a source of confusion. In video generation, PhyGround is a benchmark and evaluation framework; in human-motion analysis, it is a physics-informed force-estimation method; in adjacent work on learned simulators, “visual grounding” denotes yet another mechanism for reconciling observations with dynamics priors (Le et al., 2 Jul 2025). Precision therefore requires specifying the subfield and the cited paper.

The human-motion PhyGround method also states clear modeling limits. Its current formulation uses only global translation, assumes two-contact left/right foot prediction, and treats the body as a single mass at the root, omitting joint torques, rotational dynamics, and complex multi-contact settings such as crawling or hand contacts (Le et al., 2 Jul 2025). By contrast, the benchmark PhyGround is defined by a fixed scope of 250 annotated text+image-to-video prompts and 13 laws across three domains (Lin et al., 11 May 2026). This suggests a deliberate emphasis on observable, short-horizon physical plausibility rather than exhaustive coverage of all physical phenomena.

Across these usages, the recurring idea behind the name is not a single architecture but a methodological commitment: physical evaluation or prediction should be grounded in explicit laws, explicit observables, or explicit dynamical constraints.