Ludlow Square had two things in it: a car, and a photograph of a square. It was convincing for about eight seconds, which is roughly how long it takes to notice that nothing else in it is alive.

Putting a person on a scanned street is not the same problem as putting one in a built level, and the difference is the floor. In a built level somebody drew the floor. It is flat where it is meant to be flat, it meets the walls cleanly, and a walking character has a surface to walk on. Here the floor is a reconstruction of paving, which is to say it is uneven everywhere at exactly the scale of a footstep.

So the first walker slid, sank, and now and then stood inside a wall.

What worked was making the person a physical thing rather than an animation with a position: a body, a controller pushing it, and the same ground everything else uses. It costs more per person and it means a person can be hit by the car, which turned out to be the more interesting property and is where most of the following month went.

Then the other half of it, which is one scan colliding with another. A cactus is a scan. The street is a scan. Neither has a clean shape, both carry thousands of triangles that describe nothing in particular, and colliding them directly is both expensive and wrong: what you get is not the shape of either object, it is the shape of the noise in both.

Props collide as simple solids now, fitted to what they actually are rather than to every triangle they arrived with. It is less faithful and it behaves, which on a captured object is usually the right trade. The visible thing is still the scan; only the part you cannot see got simplified.

What I keep relearning here: the hard part of putting life on a captured street is never the behaviour. It is that everything a character stands on, walks past or bumps into was measured rather than designed, and measured things have no clean edges anywhere.