2 November 2026 · Eddie Bullock
World Model
We’re building autonomous vehicles that can scale to any city in the world. To go anywhere, our delivery robots need to know a city’s streets before they ever drive them. We're generating accurate, drivable models of real cities, inexpensively and at scale, without collecting any bespoke footage, and our robots practise every route as many times as we like before they drive it for real.
Everything on this page is the output of our first World Model pipeline. It proves we can rapidly generate high-fidelity models of real places: metre-accurate, fully labelled and checked against the real street, from London to Tokyo.
Building a simulated city, automatically
Our self-driving software is trained and tested in simulation before it reaches a public road, so the simulated world has to match the real one. We turn data for a district into a georeferenced world, placed on the map in real metres.
The videos below show each district building itself in order: roads first, then pavements, kerbs and markings, then buildings, street furniture and trees. Select a location to see the scale of one district.
Mapping our streets from laser scans
Many simulated cities are modelled by hand or generated from rules. Ours start from laser scans taken from the air, with kerbs, markings and centrelines recovered from them, so our vehicles train on the streets they will drive.
In 'Street Drive' and 'Junction', the top row shows the scan, the recovered kerbs and markings, and the finished semantic world, where every surface knows what it is. 'Real Photo' projects our world into real street photos: where a box misses its object, the error is our accountability.
Closing the sim-to-real gap
Software trained on clean simulator renders can falter on real camera footage, a problem known as the sim-to-real gap. We turn each simulated drive into realistic video while every car, kerb and lane stays put, then score the result against what we know is there. Drag across the video below to compare the simulator render with our photoreal footage.
Keeping long drives consistent
Video generators drift over long clips, inventing streets after about 20 seconds. We generate short pieces and join them, holding the layout for a full minute at the same compute cost.
Choosing settings by score
Settings within our model allow us to change how closely our photoreal footage follows the simulator's picture. We run the same drive with different settings, score each result against our world, and keep the settings that stay most faithful to it.
Below, the top left is our simulator's own render; the other three are photoreal versions made with different settings. Compare the road markings, buildings and trees in each with the render: the closer they match, the higher that version scores.
Checking our world against real photos
A simulator is only as trustworthy as its last check against reality. Each frame below is rendered at the exact spot and angle of a real street photo, with nothing pasted in. Drag across to compare: where the two line up, the world is proven right; where they don't, we know what to fix.
Ground truth without hand labelling
Driving models learn from ground truth: frames where every pixel is labelled and its distance known. On real footage, that is done by hand, frame by frame. Because we build our world, every frame arrives labelled, with depth and object edges, at no extra cost.
The videos below show our robot on a street and at a junction, a driver's view, and parked cars placed where real photos show them.
Testing in a world that reacts
Much simulation testing is open loop: recorded traffic plays out the same whatever the vehicle does. We test in a closed loop, where traffic and pedestrians react to our robot as it moves, at levels calibrated to real traffic counts. The videos below show a signalised junction and vehicles giving way at a pedestrian crossing.
Beyond the test loop: Tower Bridge
A driving system has to work beyond the streets it was developed on. We rebuilt Tower Bridge, with its traffic, crowds and river crossing, using the same pipeline as Park Royal. Our model works on landmarks in central London as well as on the roads outside Minimal HQ.
The videos below follow our robot on the road, show the bridge itself and cross it from a driver's seat.
Seeing what the driving software sees
On the road, a driving model is judged only by what it does. In our world, we can also see what it believed. In the video below, red boxes are what our software detects and green boxes are where each car really is. The red line is its planned path; the green line is the path taken.
Rebuilt streets, labelled frames and the same scoring applied to every drive, in any city in the world.
What’s next
This is v0.1, the first output of our pipeline. Next, we’re adding photorealism to every world we build, alongside huge gains in fidelity and coverage, so the worlds our robots learn in look as real as they measure.
We’ll publish updates here as we go 🤖