This study examined functions of self-motion and visual cues in updating people's actual headings in multiscale spaces. In an immersive virtual environment, the participants learned objects' locations inside two misaligned rectangular rooms by locomoting within and between the rooms. In each testing trial, the participants locomoted to adopt an actual perspective in one room, and then they judged relative direction to a target from an imagined perspective in the other room (remote perspective taking). The imagined and actual perspectives had the same or opposite cardinal directions (globally aligned or misaligned) or had the same or opposite orientations defined by room structures (locally aligned or misaligned). Global or local sensorimotor alignment effects mean that performance is better when imagined and actual perspectives were globally or locally aligned than when misaligned. We examined these effects to infer updating actual headings in global or local representations. The results showed local but no global sensorimotor alignment effect. By contrast, there were both global and local sensorimotor alignment effects when the participants judged across-room relative headings prior to remote perspective taking. These results indicate that people update headings in local representations based on visual similarities between local spaces. People update headings in global representations based on self-motion cues available in across-boundary navigation, but updating headings globally requires tasks to activate global-relevant sensorimotor representations.