Between the Flights: World Models for Physical Infrastructure

01Three kinds of world model

In 2026 the phrase world model came to cover very different systems. Google DeepMind's Genie 4 generates interactive, photoreal scenes with native sound, steered in real time by a Gemini model that reviews incoming frames to check each step succeeded ^[1]. NVIDIA's Cosmos 3 folds language, image, video, audio and action into one open model ^[2], and its 4B-parameter edge version runs robot control at 15 Hz on a Jetson module ^[3]. World Labs' Atlas reconstructs explorable 3D scenes, with explicit point clouds and Gaussian splats, from as few as two or three photographs ^[4].

World Labs also proposes a taxonomy that tells these systems apart ^[5]. A renderer outputs observations: pixels for a viewer. A simulator outputs state: geometry and dynamics that a program can compute on. A planner outputs actions. Underneath all three is the old loop of control theory, the partially observable Markov decision process: an agent acts, the state of the world changes, and the agent only ever receives observations of it. State, in their words, is “never directly visible to any agent”.

Most visible progress has been in renderers, and their limits are now measured. DeepMind writes that Genie 4 maintains coherent environments over short sessions, while visual details and object permanence “can still shift or drift over time” ^[1]. WorldProbe ran ten state-of-the-art video world models through more than 1,600 rollouts of up to 180 seconds and found that picture quality now holds for longer while “long-range memory and evolution remain weak” ^[6]. An audit of Physics-IQ, the main benchmark for physical plausibility, revised 57.6% of its samples and changed how the models rank ^[7].

02An inspection is an observation

A utility-scale solar plant has hundreds of thousands of modules, thousands of strings, and a smaller set of combiners, inverters, trackers and transformers. A thermal flight sees each module for a fraction of a second, once or twice a year. SCADA sees the plant continuously, but only at the level of inverters and combiners. Work orders and warranty claims record, in text, what crews found and did. Each is a partial, noisy view of the asset's state, taken under its own conditions.

Power loss accumulates between those views. Raptor Maps' 2026 report, covering 373 GW of US utility-scale and commercial solar, puts average power loss at 5.08%, with 4.46% already present at commissioning and a 55% rise between the second and third years of operation, as EPC warranties lapse ^[8]. Sites inspected twice in 2025 showed 7% lower loss than sites inspected once; sites inspected quarterly, 36% lower; sites using autonomous drones averaged 3% ^[8].

That is an association across sites, without a controlled test behind it. Its direction matches what operators report: most change in an asset's condition happens between observations. We therefore write the problem the way a control engineer would. Each component i carries a hidden state that evolves under weather and under the actions crews take, and each sensor k sees that state through its own observation model.

Keeping h apart from f separates the state of a component from the conditions it was observed under. A hotspot that reads 18 °C above its neighbours at 950 W/m² and 9 °C at 600 W/m² is one state seen twice. Scored flight by flight, it would read as two findings of different severity. A string that underperforms in SCADA and shows nothing in thermal imagery is evidence about the same hidden variable from a different instrument. A system that scores each flight on its own can make neither inference.

03Identity is the primitive

A state has to belong to something. Before a model can learn how string 14-B behaves, every observation of string 14-B has to be assigned to it: across flights flown at different altitudes and headings, across thermal and RGB sensors, across SCADA tags and work-order text that name it differently.

This step used to be expensive and is getting cheaper. MV-PVMapper couples SAM 3 with drone photogrammetry and, with no task-specific training, recovered 7,710 of 7,711 reference modules on three operational plants with no false positives ^[9]. Its ablations make the same point we draw below: constraints from known module dimensions separate one instance from the next. CSIRO's Tricorder links each defect found by a ground robot under the array to a specific asset and its 3D position ^[10]. GS-Pool finds which objects changed between two separate visits and returns, for each, the evidence an inspector can review in 3D ^[11]; ChronoFuseGS merges captures from different days into one model in which every element records when it was seen ^[12].

Our approach treats identity as constrained assignment. The plant's layout, from as-built drawings or a first survey, defines the set of components that exist; every detection is assigned to one of them and flagged exact, conflict or approximate, and conflicts go to a person. Free matching between flights can merge or invent components without anyone noticing. Assignment to a known lattice cannot do either silently.

Generative reconstruction carries a further risk. World Labs is candid that Atlas fills regions no camera saw: “the more it sees, the less it imagines” ^[4]. In an inspection record, an imagined module is a false record. Anything a model infers without observing it has to be marked as inferred and kept out of the evidence.

04Predict the state

If the state is hidden, a model has to predict it in some representation. The 2026 evidence favours latent prediction over pixel prediction. A matched comparison of four frontier video encoders found that the latent-prediction models “degrade more gracefully under pixel corruption”, keep class structure under occlusion and uniquely encode the direction of time; only latent prediction gave near-complete task success when paired with a planner ^[13]. Meta's V-JEPA 2.1 produces dense, temporally consistent features of the kind that tracking parts of a scene needs ^[14]. RoboJEPA scaled a latent predictor to 8B parameters across twelve robot types and found that the error of its imagined rollouts follows a power law in compute and is “a reliable proxy for real-robot evaluation” ^[15].

Three further results map almost directly onto assets.

• –Timescales. In H-JEPA, when factors evolve at separate timescales, higher levels “discard fast, unpredictable detail and retain slower task-relevant state” ^[16]. On a solar plant, irradiance changes in minutes, soiling over weeks and degradation over years. The state should carry the slow processes; the observation model should explain the fast one.
• –Physics in training only. JEPA-x uses physical state as a privileged signal during training and deploys on images alone, cutting rollout drift from 0.361 to 0.104 ^[17]. The analogue for an asset is training with IV curves, SCADA and post-repair measurements, then running where only imagery exists.
• –Surprise as a signal. LeWorldModel, at about 15M parameters, detects physically implausible events from its own prediction error ^[18]; EpicWorldModel uses the variance of its predictions to decide where to look next ^[19]. For an asset, prediction error marks a candidate finding, and predictive variance marks where the next flight should look.

The closest published work to our formulation is GTSM, which represents thermal evolution as a shared scene-level latent process plus a local thermal state for each Gaussian, and predicts thermal state beyond the observed interval ^[20]. Replace the Gaussian with a module and the scene with a plant, and it is equations (1) and (2).

05Physics as a guardrail

Physics does not improve every model. In the largest test this year, across 3,713 PV installations in 33 European countries, physics-informed hybrids showed no consistent improvement in accuracy, generalisation or performance with little training data ^[21]. Physics helps in narrower settings. A physics model with learned residuals cut day-ahead error by 38.5% against physics alone and kept predictions feasible under corrupted measurements ^[22]. Physics-generated histories let forecasting models work at a new plant before it has produced any data ^[23]. A Bayesian physics-informed model of transformer insulation ageing, validated on field data from a solar plant, separates uncertainty that comes from the world from uncertainty that comes from the model ^[24].

Our position is to use physics in that narrower way: as a constraint where data is thin and as a guardrail when a sensor fails. A model that falls back to a plausible physical answer when a string monitor drops out is worth more on site than a slightly more accurate one that returns physically impossible values.

06Test forward in time, at sites the model has not seen

Most results in this field are reported on random splits of images. For an asset model that is the wrong test: it lets the model see the same site, often the same afternoon, in training and in test. The 2026 literature shows how much that hides.

• –Scored on future data as it arrives, the rankings of time-series foundation models “undergo a dramatic reshuffling” across eleven domains ^[25].
• –On 546 fixed camera sites followed over time, foundation models underperform at many sites, and naively updating them on past data can push them below zero-shot ^[26].
• –Under a real, measured shift in operating conditions, 95% conformal intervals on a thermal model covered 70.13% of outcomes; reweighting recovered only 72.42% ^[27].
• –In drone power-line inspection, once image resolution and conditions were matched, no vision-language model family kept an advantage over a plain ResNet-50 after correcting for multiple comparisons ^[28].
• –A model pretrained on a PV fleet beat generalist foundation models at 139 stations held out by location ^[29].

The protocol we propose follows from these results. A model is trained on some sites up to a cut-off date and tested on other sites after it, with a gap between the two so that weather and season cannot leak across. Uncertainty is judged by its coverage at each site. Baselines are mandatory and deliberately simple: carry the last observation forward, apply the rules a good O&M engineer would, run the physics model alone. The final score is the recorded outcome: whether the predicted fault became a repair, what the repair cost, and whether the fault came back.

07What we are building

Three things are now cheap that were not two years ago: module mapping from drone imagery ^[9]; latent predictive models small enough to train for one fleet ^[18]; and calibrated forecasting that runs on a CPU ^[30], next to world models that run on an edge module ^[3]. That moves the bottleneck to the record that connects a finding to a repair, to its cost, and to whether the fault returned. We have not found a public dataset that links those four.

Aperture builds that record today across six utility-scale solar parks in Hungary: 82,368 radiometric thermal frames from seven campaigns between 2023 and 2025, and 2026 campaigns flown at all seven site presets. The largest layout holds 116,391 modules. Every detection is assigned to a module of the surveyed layout and flagged exact, conflict or approximate. Every finding is audited by an expert before it is published, with one of nine causes and one of eight next steps, and verifying it opens a follow-up that the next campaign reads, treated or not.

Each step beyond that runs behind a gate. Module-level history across campaigns switches on when the wrong-row rate is at most 1.0% at 90% coverage or more; until then, history is held per string. A next-condition forecast ships only when it beats the last observation forward in time; until then, we report the last observation. Equations (1) and (2) define two targets: the state of each component today, given everything observed so far, and the action that changes its next six months the most. The value of that action is a counterfactual, reported with its method and interval where diagnostics pass, and elsewhere as bounds, or not at all.