A learned walking model, on experiments and days nobody had opened

Living Worlds' crowd, with a learned layer on how people walk, 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, on each of two seeds: 25 of 29 runs of a Jülich experiment with open entry, 8 of 9 runs of a second Jülich experiment, and 11 of 12 three-minute windows on two days in a shopping centre in Osaka. The bar was 20, 6 and 8. A run where a crowd failed to finish counts as a loss: one in the first set and one window in the third, on both seeds. The protocol, the code and the settings were frozen and checked by hash before any of the data was opened, and the test ran once. What it shows: on people walking the same way, in a corridor and across a concourse, the learned layer is closer to the real measures of walking speed, spacing, path straightness and close contacts than the crowds it was set against. It also ranked ahead of straight walkers in 25, 9 and 11 runs and ahead of ORCA in 27, 9 and 11. What it does not show: anything about counterflow, crossings or bottlenecks, where the packed-crowd time gap in the layer was not tested; the crowd knows where and when each person really entered and left, as the straight walkers do; the two seeds rerun the same model on the same recordings, so they show that the result does not hang on one random draw, not that it holds on other data; and the time gap was chosen after seeing two of the three development seeds, with thin margins in development.