How many robots does it take to change how a crowd walks?
research/society-robotnorms
How many robots does it take to change which side a crowd walks on? Robots that hold a programmed side tip a crowd once they are 16–32% of it, more when people remember more past partners (simulation matches mean field within 0.02); just short of that they make people collide up to 14× as often as adaptive robots, and placement matters more than numbers when people meet only neighbours. Stylised.
What it found
Robots as a committed minority in a passing-side convention. Simulated people take the side most of their last M partners used; robots hold a programmed side. The tipping share grows with memory: 16%, 22%, 28% and 32% of the crowd for M = 5, 9, 17 and 25, within 0.02 of a mean-field prediction. Just below it, robots that hold their side make people collide up to 14× as often as robots that adapt (0.27 vs 0.020 conflicts per meeting). When people meet only neighbours, evenly spread robots tip everyone from 12–14%, while one block of robots converts under 10% of people even at half the crowd. Stylised: one binary choice per meeting, no geometry.
A figure from its own output
Paper, code and results
More in Robot societies among people
- Do robots that keep to one side make people walk on it, or jam the corridor?Do robots that keep to one side make people walk on it, or jam the corridor? In a social-force corridor, robots that adapt change nothing, but robots that hold the left cut people's speed from 0.91 to 0.39 of desired by 30% of the crowd and halve lane separation, while people move to their side only gradually (no sharp tipping, unlike the 1D model). Stylised.
- Can a robot fleet's reputation survive one bad incident?Can a robot fleet's reputation survive one bad incident? In a model where trust spreads by experience and gossip, a dent always recovers because people keep meeting robots; reputation is lost only through avoidance (at a 0.30 avoid-below threshold, 20/20, 18/20 and 0/20 crowds recover under neutral, 2× and 4× negativity in gossip), and robots that explain failures recovered every time. Stylised.
- What does it cost robots to let people go first at a shared lift or narrow pavement?What does it cost robots to let people go first at a shared lift or narrow pavement? At load 0.9, deference cuts people's wait 6.5× (4.60 to 0.71) and raises robots' by 2.6 units; total waiting cost is conserved, so it pays only if a person's wait costs more than a robot's, and a lone defector waits far less (2.08 vs 7.65). Queueing model within 2.7% of closed forms.
- Do robot agents keep the norms they state?Do robot agents keep the norms they state? A pluggable town harness with a deterministic stub (no language model run): a robot stating "always yield" keeps it in 1 − 0.4×pressure of meetings (1.00 to 0.60), and people's cost per meeting rises from 0.20 to 1.08. Stylised.
- Does the robots-in-a-corridor conclusion survive a more humanlike crowd?Does the robots-in-a-corridor conclusion survive a more humanlike crowd? Adaptive robots stay free and committed robots never flip the crowd in all five simulated crowds, but the size of the slowdown (loss of 0.17–0.49 of desired speed at 30% robots) and of personal-space intrusions (0.03–0.36 of agent time) depends strongly on how humanlike the people are. Stylised.
- How many robots does it take to change a crowd's walking side when some people never adapt?How many robots does it take to change a crowd's walking side when some people never adapt? With 0, 20, 40 and 50% stubborn walkers the mean-field tipping share (memory 9) rises 0.217, 0.240, 0.283, 0.329, simulation agrees within 0.02, and committed-robot conflict costs barely change. Stylised.
All 31 Robot societies among people projects →
Other research topics
Proper scoring & elicitation28Effort & peer prediction9Verification games & slashing19Detection & tolerances14Markets as routers13Prediction-market design9Paying contributors8Decentralised training systems23Digital twins & sim-to-real164LLM chip design loops40Robot manufacturing loops53Ambient AI assistant46Self-replicating manufacturing34Simulation-first design64Engineered clothing45Task-specific AI accelerators40Algorithmic chip-design loops42Open-source purpose-built robots44The software-engineering loop43Adversarial human imitation41Open-source residential construction41Autonomous multi-angle capture34Sensing, imaging and 3D31Energy supply and dense energy storage39Robot societies among people31Agentic co-design of low-cost robots27Laundry-hanging robots25Machine learning on qualitative data23Robots on the building site27Civilizations & group psychology1History & human sciences1
Produced with AI assistance (Claude). Results are from the code in this repository and are simulated or stylised unless the abstract says otherwise; nothing here implies affiliation with or endorsement by any lab, researcher or company cited. MIT licensed.