The brief
Every routing system in common use answers the same question: what is the shortest path from here to there. It is a good question for a delivery van and a poor one for a walk. Two kilometres along an arterial road and two kilometres through a park are identical to a router and nothing like each other to a person, and that difference is most of the reason anybody walks for pleasure at all.
Walking is not only transport. It carries physical health, mental wellbeing, social contact and attachment to place, and planners increasingly accept that a walkable city needs more than pavements that connect. It needs walks worth taking. But the qualities that make a route worth taking are exactly the ones routing engines discard, because they are qualitative, contested and hard to put in a cost function.
WalkGrid exists to put them in the cost function anyway. The wager is that enough of what makes a place good to walk through can be measured, and that somebody who wants a green quiet walk with a couple of listed buildings on it should be able to ask for that and get it.
The insight
The organising idea is environmental similarity: two places offering comparable experiences will have comparable environmental profiles. Tile the city, score every tile across enough dimensions, and “find me somewhere like this” becomes a distance computation in a 51-dimensional space. The argument for treating walking context this way is set out in a spatial-platial vision paper, and the recommendation model it produced in an emerging conceptual model.
That reframing does real work. Route curation stops being an optimisation problem with one right answer and becomes a search problem with many defensible ones, ranked by a preference the user actually stated. It also makes the system honest. Because the weights are explicit and the features are named, you can always ask why a route was proposed and get an answer.
The harder discovery came from evaluation. The V1 study confirmed the model and condemned the interface. People do not think in fifty-one features. They had already told us as much in a survey of leisure walking behaviours and in the platial narratives that came out of the walking study. They think in phrases: somewhere green, not too busy, interesting to look at. Asking them to decompose a phrase into weighted sliders imposed exactly the cognitive work the system was supposed to remove. The abstraction gap was not a usability detail. It was the thing standing between a working model and a usable tool.
V2 puts a language model in that gap. The user says the phrase, the model performs the decomposition. Crucially the mapping stays visible, because the selected features and inferred weights are shown and can be corrected, so the model acts as a translator rather than an oracle.
What it does not do
It inherits the quality and the currency of its sources. A crime statistic and an air quality reading age at different rates, and the grid does not currently distinguish between them. It cannot measure the things people most often name: quietness, beauty and feeling safe are proxied by lighting, CCTV and road density, which is not the same thing and should not be mistaken for it.
It has no detailed accessibility layer. There are no kerb cuts, gradients are rendered only as elevation variance, and there is no step-free guarantee, which limits its usefulness for precisely the walkers with the least margin for a bad route. It also covers one city. Extending to another is a data pipeline problem rather than a research one, but it is not free.
The equity risk is worth stating plainly. A system that surfaces the most pleasant environments will, without care, route people away from places that are already under-invested in, and reinforce the pattern it is measuring.
The full argument, including the evaluation studies in detail, is in the thesis: A Framework for Curating Personalised Leisure Walking Experiences. The wider standards problem this work kept running into became its own project, the Leisure Walking Systems Working Group.
