Blog · Research · Oct 21, 2026 · 5 min
Living Worlds' crowd against real pedestrians in Jülich experiments, a shopping centre in Osaka and six recorded street scenes, each scored once on recordings nobody had opened: met in the shopping centre, in four fresh Jülich experiments and with a learned walking layer; not met in the first Jülich set and in three of six street scenes.
When a crowd moves through a corridor, a crossing or a shopping centre, how fast do people walk, how much space do they keep, how straight are their paths, and how often do they come close? A simulated crowd that gets these right can stand in for real people when a station, a hall or a street is planned. One that gets them wrong carries its errors into every decision built on it.
Living Worlds asks whether its crowd moves more like real pedestrians than the standard models do. The comparison is social force, the long-standing pedestrian model, and people who ignore each other and walk straight, with ORCA scored alongside in later tests.
Every crowd test here was registered before the recordings it was scored on were opened, and scored once. The crowd knows where and when each real person entered and left, as the straight walkers do; everything in between, the path and the pace, it has to produce. It is then scored on four measures from the real tracks: walking speed, distance to the nearest person, path straightness and how often people came within 0.3 metres.
Opening a recording before the rule is set invites quiet tuning towards it. Keeping it closed means a win is a prediction, and a loss is recorded as a loss. Breadth matters as much: a 2025 study notes that current data-driven crowd models are primarily designed for a single scenario, with very few validated in more than two. The tests below span counterflow, crossings, corners, bottlenecks, shopping concourses and city streets.
The Jülich experiments are controlled recordings of volunteers walking through set layouts. Crowds in four Jülich experiments scored 109 runs across counterflow, a right-angle crossing, merging at a T-junction and a bottleneck with groups. With groups added, the crowd beat social force and the straight walkers in every run at the T-junction (11 of 11), but not in more than half of counterflow (3 of 15), the crossing (10 of 43) or the bottleneck (17 of 40). Not met (1 of 4).
The first pass stopped on a bug before anything was scored; the protocol was amended to fix only that, and every run was re-run once. A correction followed on 11 October 2026: 19 of the 43 crossing files state metres in their own header but were read as centimetres. The count includes them and is left as registered.
Crowds against real pedestrians, six recorded scenes used the public ETH and UCY scenes, none run before. Living Worlds' crowd moved more like real pedestrians than both social force and the straight walkers in 3 of 6 scenes, where 5 were needed. Not met (3 of 6). It won the three quieter scenes. In the dense ones, two at a university and one outside a shop, real people walk in groups and the crowd kept everyone apart.
A new crowd followed. Crowds in four Jülich experiments nobody had opened scored 52 runs, with ORCA alongside, and any run where a crowd failed to finish counted as a loss. It moved more like real people than social force and the straight walkers in more than half the runs of all four experiments: the 120-degree crossing 9 of 16, counterflow 11 of 18, the corner 8 of 13 and the bottleneck 4 of 5. Met.
A corridor tells the crowd a great deal: people go in at one end and out at the other. Shoppers wander, so where they enter and leave says little about their path. Crowds among shoppers, a day nobody had opened took the same crowd, unchanged, into the ATC shopping centre in Osaka, where shoppers are tracked by range sensors. It beat both social force and the straight walkers on all four measures in all six three-minute windows of the day, where five were needed. Met (6 of 6).
A learned walking model, on experiments and days nobody had opened added a learned layer on how people walk. The protocol, the code and the settings were frozen and checked by hash before any data was opened, and the test ran once. On each of two seeds it moved more like real pedestrians than both the earlier Living Worlds crowd and social force in at least two thirds of the runs of every set: 25 of 29 runs of a Jülich experiment with open entry, 8 of 9 runs of a second Jülich experiment, and 11 of 12 windows on two days in the Osaka shopping centre, against bars of 20, 6 and 8. Met (3 of 3 sets, on both seeds). It also ranked ahead of straight walkers in 25, 9 and 11 runs and ahead of ORCA in 27, 9 and 11. A run where a crowd failed to finish counted as a loss: one in the first set and one window in the third, on both seeds.
The record shows a crowd that got better in public. The first Jülich set and the street scenes were published as not met, with the reasons stated: counterflow, crossings and bottlenecks were hard, and dense street scenes need people who walk together. The next registered tests met on four fresh Jülich experiments, a full shopping-centre day and, with the learned layer, on two more experiments and two more days. A 2025 scientometric analysis of data-driven pedestrian work describes a strong convergence between artificial intelligence, urban informatics and crowd behaviour modelling. What these results add is a register of single scored runs, wins and losses alike.
The learned layer was tested on people walking the same way, in a corridor and across a concourse. It shows nothing about counterflow, crossings or bottlenecks, where its packed-crowd time gap was not tested. Its two seeds rerun the same model on the same recordings, so they show the result does not hang on one random draw, not that it holds on other data, and that time gap was chosen after seeing two of the three development seeds, with thin margins in development. In every test the crowd knows where and when each real person entered and left. And dense street scenes with walking groups remain the open case: 3 of 6 is the recorded result.
Living Worlds simulates the places people live in as well as the people: how a room sounds, how light falls, how crowds move, how power flows. Each is checked against real measurements before it is relied on, and used within the range where it held. For crowds, that range is now set by these registered tests, met and not met.
Sources
Blog
Physical AI now runs on open physics engines, world foundation models and synthetic demonstrations measured in hundreds of thousands. The newest benchmarks ask whether simulated scores predict real ones, and the part of the scene least checked against reality is the people.
An agent that remembers what a person said predicted their answers four years later better than a persona, in every panel and on every question; an agent given only a life record did no better than a persona. And Living Worlds' crowd predicted how close real people came to a real robot in 7 of 8 recordings nobody had opened.
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.