The robots that seemed to speed up a crowd were mostly just making room

Crowd-side robots appeared to cut walkers' transit by 1.6 s. With a control that keeps the people count but drops the robots, 88% of the cut is density.

The last step of the society-lanes programme found something it had not been looking for. In a simulated corridor where people's keep-right habit was weakened, robots that held the crowd's side seemed to make people walk faster: 1.6 s off each person's trip through an 8 m section, with people's lanes looking tidier too. If robots can act as lane seeds, that is a reason to deploy them beyond "they brush past fewer people".

That reading had an obvious problem. In those runs a robot replaced a person, so a corridor with 20% robots also held 20% fewer people. Fewer people walk faster. So this step registered six predictions and ran a control that only takes the people out: the same number of people as the robot runs, and no robots.

What was run

The social-force crowd model in JuPedSim, fitted earlier to real Jülich corridor experiments, in two of those corridors (30 and 60 agents). Four conditions: the full crowd, the thinned crowd (the control), the thinned crowd plus robots on the crowd's side, and plus robots on the other side. Robots at 10% and 20% of the agents. The keep-right bias was swept from a quarter of its fitted value to the full value. Twenty fresh random seeds per cell, 1,120 runs, intervals from a bootstrap over seeds.

What happened

At half the bias with 20% robots, the setting of the original hint:

Change in people's transitSeconds95% interval
Robots + fewer people (total)−0.91[−1.58, −0.11]
Fewer people alone (density)−0.80[−1.37, −0.26]
Robots beyond the control (seeding)−0.11[−0.74, +0.66]

Density accounts for 88% of the speed-up. In none of the eight bias and share settings did crowd-side robots shorten people's trips beyond the control with an interval clear of zero. The earlier 1.6 s was itself one draw of 20 seeds; this draw gave 0.9 s at the same setting.

Two predictions held. Robots on the wrong side slow people by 1.2 to 4.3 s against the same control in every setting. And at a quarter of the fitted bias the crowd kept right 68% of the time, the first simulated crowd in the programme to sit at the value measured on real city pavements (65% in Zurich, 68% in Nicosia).

One thing showed up that was not registered: at 20% robots, crowd-side robots did raise people's lane order a little beyond the control (+0.08 at a quarter of the bias, +0.055 at three quarters). Tidier lanes, then, but not faster walking. That needs its own registered test before it counts.

What it means

A fleet operator choosing a passing policy still has the reason from the last step: holding the crowd's side gives the fewest close passes, and three different crowd models agree on that ranking. "Robots make the crowd faster" is not a reason; it was the crowd getting thinner. The caveats stay: one crowd model family, simulated people, robots that people can push.

Code, results and the registered predictions: research/society-lanes (experiments/m6c_lane_seeds.py, LOG.md).

All writing · More in Research findings · More on society-lanes · All tools and research