Yielding robots hit 74–77% on Penrose tiling

The experiment sits on a Penrose rhombus floor whose 90 vertices are addresses; the left 32 form a retina, where an average of 2.8 addresses light up. Each exit has a scent layer computed by adding neighbors' values, and robots step toward the strongest neighboring scent. The core change is stacking yielding rules on top — one robot per address at a time, sidestep if the strongest neighbor is taken, wait if both are blocked, and switch layers if a pass is standing at the blocked spot.

What the results show is that coordination rules do more than prevent jams. When passes are removed, robots overlap and slip through, turning the task into a simple majority vote per address, and accuracy drops from 73.7–77.2% to 61.7–64.2%. The author also found that learning layer-switching passes produced 14 passes split by orientation: some work only for vertical bars while others fire only for horizontal ones. Waiting itself differs by orientation, measured at 2.03 times per bar for vertical and 1.61 for horizontal.

The author frames this as a companion to local-rule research in swarm robotics, not a new method. The test ahead is whether the same approach holds when the answer has three or more categories, or when the floor and image scale up.