Predictable movement matters as much as safe movement
A movement can be perfectly safe in the physics sense, no risk of contact, adequate clearance, and still be alarming if the person nearby can't predict it before it happens. This is because humans navigating shared space rely heavily on anticipation: when walking near another person, we constantly extrapolate their trajectory a few steps ahead and adjust our own path accordingly, usually without conscious thought. Robots that move in ways that resist this kind of extrapolation, sudden accelerations, unflagged direction changes, silent starts after being stationary, force people to abandon their normal anticipatory strategy and instead react defensively after the fact. That reactive posture is what reads as discomfort or alarm, independent of whether the robot was ever actually unsafe.
Legibility is the design property that closes this gap: a robot is legible when a nearby human can correctly infer its near-term intent before it acts, using the same kind of cues they'd use to read another person's intent. This is a genuinely different target than motion planning accuracy. A path planner cares whether the trajectory is efficient and collision-free. Legibility asks whether an observer watching the robot's first half-second of movement can already guess where it's going. A trajectory can be optimal by every planning metric and still be illegible if its early motion looks the same whether the robot is about to turn left or continue straight, leaving the observer no signal to act on until it's too late to smoothly adjust.