The MapScaping Podcast - GIS, Geospatial, Remote Sensing, earth observation and digital geography

In this episode I'm joined by Apurva Shah, co-founder and CEO of Duality AI, a company building virtual worlds — or "world models" for robots and physical AI systems.

Apurva's path here is an unusual one. He spent most of his career in animation, first at Pacific Data Images (which later became DreamWorks) and then over a decade at Pixar. His co-founder, Mike Taylor, comes from the other end of the spectrum entirely: a controls engineer who led field robotics at Caterpillar, deploying house-sized haul trucks at Australian mines. As Apurva puts it, if he's the pixels, Mike is the atoms.

We talk about why real-world data, as valuable as it is, is never enough on its own — and how synthetic data can be used to deliberately fill the gaps and biases that creep into any collected dataset.

Some of the things we get into:

The difference between digital twins and 3D assets and why Duality treats twins as modular building blocks you compose into scenarios, rather than as one monolithic environment

How they build environments from the ground up using DEM data, satellite imagery, photogrammetry and biome catalogues and why building them this way means everything is annotated from the start

Calibrating virtual sensors against real ones, including synthetic aperture radar, and why sensor noise characteristics matter as much as physics

Predicting how a material will behave across the spectrum (infrared, SAR) just from its visual response — and when that prediction breaks down

Why "clutter" only becomes clutter once you know what you're looking for, and why it doesn't need to be perfect

Modelling star fields for localisation in space, where there are no roads or buildings to navigate by

Explicit versus generative world models, and why you need both

A project with AWS simulating emergency ambulance routing through a city, complete with autonomous vehicles, traffic control and teleoperated human agents

Where Duality is not the right tool molecular scale, virtual patients, drug discovery

And yes, a story about robotics companies renting Airbnbs, trashing them, and leaving

Towards the end we get into the bigger questions: whether AI takes our jobs or makes us better at them, where the line sits between "good enough" and slop, and why Apurva — a self-described humanist — thinks virtual environments are the one place where human and machine intelligence can genuinely learn from each other.

Find out more at duality.ai, or dig into their technical writing at duality.ai/blogs

My guest today is Tyler Reid, co-founder and CTO of Xona, a company building the first commercial satellite navigation system.

We get into why Tyler and his team are moving satellites into low Earth orbit and what that unlocks. Stronger signals that can penetrate indoors, more resilient timing infrastructure, and better security against jamming and spoofing.

GPS sits 20,000 kilometres out. Xona sits at 1,100, with signals around 100 times stronger and a planned constellation of 258 satellites. Tyler came at this from the autonomous vehicle world at Ford, where the problem was simple enough: ten meters gets you to the store, but it doesn't keep a car in its lane.

We also talk about time, and how much of the world quietly depends on GPS to keep its clocks honest. Why countries are suddenly so interested in owning their own infrastructure. And the question Xona gets asked constantly: if you're broadcasting that close to GPS, aren't you the jamming problem?

If you're interested in what the future of GNSS might look like, you're really going to enjoy this one.

More at xonaspace.com, or reach out to Tyler on LinkedIn.

Vexcel isn't a household name — but you've almost certainly used their data. This aerial imaging company flies low-elevation aircraft across roughly 45 countries, capturing imagery at 7.5cm resolution from five different angles (straight down plus four oblique views), building one of the richest geospatial datasets on Earth.

In this episode, Daniel talks with Steve Lombardi, VP of Product at Vexcel, about what happens after the pixels are captured. They dig into object detection and "elements" (pre-extracted features like roof condition, solar panels, and pools), then go deep on Vexcel's newest capability: vector embeddings — essentially a searchable fingerprint for every 100-meter chunk of the planet.

Steve explains how customers can search the visible world with a text phrase, an uploaded image, or by simply drawing a box on the map — and get back matching locations anywhere on Earth. They cover how oblique imagery adds context to searches (like finding buildings that "look like a palace"), how customers refine results with a simple thumbs up/down feedback loop, and a fascinating new use case: exposing embeddings as a QGIS tile layer so you can build a personalized, concept-driven heat map — like a custom risk map — without ever touching a database.

Topics covered:

What makes Vexcel's aerial imagery different from satellite imagery

How photogrammetric data enables precise 3D measurement and object detection

Object detection vs. vector embeddings — when to use which

Custom elements: letting customers define their own objects to detect

Searching aerial imagery by text, image, or drawn area

Refining search results with a lightweight classifier ("thumb up / thumb down")

Change detection over time (the Austin, Texas example)

Bringing embeddings into QGIS as a personalized, concept-based tile layer

Where aerial imagery and geospatial AI are headed next

What happens when you take cloud native geospatial out of the satellite-and-petabyte world and drop it into a small city with no budget and the world's slowest internet connection?

In this episode, Daniel talks with Nissim Lebovits, a city planner turned geospatial data scientist, about his nine months in Argentina building climate risk tools for under-resourced municipalities. Nissim breaks down what "cloud native" actually means in practice (hint: it's really about ease of access), why range requests let a small city grab 100MB instead of downloading a 20GB file, and how a single spatial join — run in about three seconds — revealed that 3 million people are missing from Argentina's own census data.

We also get into open building footprints, why QGIS (not Python) is where the real adoption is happening, the adoption gap holding cloud native geo back from small cities in the Global South, vibe-coding a QGIS plugin to finally make Argentina's census data usable, and where cloud native geo is headed over the next five years.

This episode is brought to you by the Cloud Native Geospatial Forum. CNG Forum 2026 runs October 6-9 at Snowbird, Utah — three days of real-world cloud native geospatial (STAC, COGs, GeoParquet, Zarr, and more) plus a hands-on workshop day on the 6th. Register at 2026.cloudnativegeo.org.

Jeffrey Martin is the co-founder and CEO of Mosaic, a company building 360-degree camera systems designed specifically for mapping. He's been obsessed with 360 imagery for over 20 years — he built one of the first websites combining panoramic images with a map back in 2005, before Google Street View existed, and he holds a Guinness World Record for a 320-gigapixel image of London stitched together from 52,000 photos.

In this episode, we get into what a modern mapping-grade 360 camera actually looks like, and why the difference between rolling shutter and global shutter sensors matters if you care about things like colorizing point clouds or photogrammetry. We also cover the surprisingly long tail of people who need up-to-date street-level imagery — everyone from departments of transport and utilities companies to playground designers and outdoor furniture salespeople.

A few things that stood out:

Ground-level 360 capture fills gaps that drones and satellites simply can't — occlusion, permissions, and viewing angle all favor eye-level imagery for certain infrastructure work.

Companies are taking very different bets on data collection strategy — Mosaic focuses on high-quality, purpose-built capture, while others like Hive Mapper are betting on scale and crowdsourced coverage instead.

The camera hardware race may be plateauing, but the real frontier now is what we do with the imagery once it's collected — especially as large language models start being layered on top of geospatial data.

Where does this all go next? If AI can eventually parse and reason over an entire country's worth of street-level imagery, what does that unlock — and who ends up owning that layer of the map?