World Models or Bust? Open-Sourcing a JEPA Model for Building Entrance Prediction

Ed Parsons recently argued that AI world models represent an existential disruption to the geospatial industry. His thesis is provocative: the entire edifice of precision measurement, authoritative databases and deterministic spatial queries may be rendered obsolete by systems that simulate reality rather than catalog it. The traditional geospatial stack, he warns, faces the same fate as encyclopedia publishers when Wikipedia arrived.

He’s right about the direction, but the timeline and mechanism might be more nuanced. At Zephr, we’ve been working at the boundary between these two paradigms. To that end we just open-sourced a model we think illustrates how the bridge between deterministic geospatial and world model reasoning could be built in practice.

The model is a Joint Embedding Predictive Architecture, or JEPA, trained to predict building entrance locations from street-level imagery and building geometry. It is available at github.com/zephr-xyz/jepa-entrance.

What JEPA does differently

A traditional approach to finding building entrances would be object detection: train a neural network to recognize doors in images, draw bounding boxes, project to map coordinates. This is fundamentally a pattern matching task. The model learns what doors look like and finds them.

Our JEPA model does something conceptually different. Instead of detecting objects, it learns to predict latent geometric representations. The context encoder observes street-level images and produces a compressed representation of what it sees. The target encoder observes the building geometry and the true entrance position and produces a compressed representation of the spatial reality. The predictor then learns to map from the visual representation to the geometric one, bridging what the camera sees with where the entrance actually is.

This is a world model in miniature. The system learns an internal representation of how visual appearance relates to physical spatial structure. It doesn’t memorize entrance locations or template-match door shapes. It learns the relationship between facades and access points.

Why world models work for entrances

Ed frames world models primarily as simulation engines for dynamic scenarios: hurricane response, traffic flow, urban planning. Alternatively there is a less obvious and more immediately practical application of world model reasoning, which is inferring spatial states that are ambiguous, occluded, or context-dependent.

Building entrances are a perfect example. They are not static properties of geometry. An entrance is a function of the building’s relationship to the street, to pedestrian flow, to architectural convention, to the surrounding built environment. The same building footprint in a strip mall, a downtown block, and a suburban office park will have its entrance in a different location relative to the facade, even if the footprint is identical. A detection model can only find what it sees. A world model can predict what it expects based on learned spatial relationships.

This is exactly where JEPA’s predictive architecture excels. The model doesn’t need to see a door. It observes the visual character of the facade, the relationship between the camera and the building geometry, and the spatial context, thenit predicts where along the facade the entrance should be. On RTK-precision ground truth data collected in Boulder and Louisville, Colorado, the JEPA model achieves a median error of .78 meters, a 69.2% improvement over geometric baselines that simply predict the facade midpoint.

JEPA Entrance Prediction Performance

It is also helpful to see the model in action with this small demo we made. For a subset of POIs in Boulder County we built out the ray tracing for each camera position to show how the JEPA embeddings are used to narrow the location of the photo with the predicted entrance. When you click on one of the Mapillary images in the right hand vertical carousel it will light up the rays in blue that align with the JEPA predicted entrance from the photo. You’ll notie the photo also has a blue transparent column for where the model thinks the entrance is. The mode isn’t perfect, but in aggregate it results in a better entrance prediction.

JEPA Entrance Prediction Model Animated

The key insight is that world models are not just useful for simulating futures. They are useful for inferring presents that are difficult to observe directly. Entrance location is a spatial state that depends on context, convention, and physics. A model that learns internal representations of these relationships will outperform one that tries to detect surface-level visual features.

Bridging deterministic and probabilistic geospatial

I don’t think we will live with binary options: deterministic databases versus probabilistic simulations. In practice, the near-term future is a hybrid. The JEPA entrance prediction pipeline take a fist step at illustrating this hybrid approach.

The building footprints come from Overture Maps, a deterministic geospatial dataset. The street-level imagery comes from Mapillary, a crowdsourced observational dataset. The camera geometry, the ray-facade intersections, the coordinate transforms are all classical computational geometry, exact and reproducible. The JEPA model sits on top of this deterministic foundation and adds learned spatial reasoning. It doesn’t replace the building polygons or the camera models. It interprets them.

We can think of the next evolution of the JEPA model being integrated with cross view embeddings. You don’t have street view imagery of the building but a cross view embedding allows you to still run your JEPA prediction. The world model approach allows you to simulate what is not there so you can still make geometric augmentations to your foundational data.

This is the pattern we expect to see across the geospatial industry over the next few years. World models will not arrive as monolithic replacements for existing spatial data infrastructure. They will arrive as reasoning layers that sit on top of authoritative geometry and add inference capabilities that deterministic systems cannot provide.

Why we’re open-sourcing this

The JEPA entrance model is small: it trains in seven minutes on a single GPU. The dataset is modest: roughly 1,200 samples. The problem is narrow: predicting a single parameter along a building facade. But it demonstrates a principle that we think matters. World models are not a future abstraction. They are a practical tool that can solve real geospatial problems today, right now, when combined with good spatial data and precise positioning.

We’re releasing the model, the training code, and the evaluation pipeline because we think the geospatial community needs concrete examples of what the bridge between deterministic and learned spatial reasoning looks like. Often the conversation about world models in geospatial is speculative. This is working code with measurable results against centimeter-precision ground truth.

The model is at github.com/zephr-xyz/jepa-entrance. We welcome contributions, critiques, and extensions.

Where this leads

Ed asks the right question: will world models remake geospatial? The answer is yes, but maybe not yet in the way the most dramatic predictions suggest. The next step may be augmenting models that understand spatial relationships the map cannot encode.

Entrance prediction is a small example of a large pattern. Every spatial property that depends on context, convention or physics rather than pure geometry is a candidate for world model reasoning. Pedestrian flow patterns. Optimal delivery drop-off points. Accessible routing through complex buildings. Sight lines and wayfinding legibility. These are all spatial inference problems where the right answer depends on understanding how the built environment works, not just what it looks like from above.

The geospatial industry has spent decades getting extraordinarily good at measuring the world. The next chapter is about understanding it. World models are how we get there, and we think the path runs through practical, grounded applications like this one. The open question is are these just stops along the way to full scale displacement by simulation centered models.