One of the key questions in spatial perception is whether the brain has a common representation of gravity that is generally accessible for various perceptual orientation tasks. To evaluate this idea, we compared the ability of six tilted subjects to indicate earth-centric directions in the dark with a visual and an oculomotor paradigm and to estimate their body tilt relative to gravity. Subjective earth-horizontal and -vertical data were collected, either by adjusting a visual line or by making saccades, at 37 roll-tilt angles across the entire range. These spatial perception responses and the associated body-tilt estimates were subjected to a principal-component analysis to describe their tilt dependence. This analysis allowed us to separate systematic and random errors in performance, to disentangle the effects of task (horizontal vs. vertical) and paradigm (visual vs. oculomotor) in the space-perception data, and to compare the veridicality of space perception and the sense of self-tilt. In all spatial-orientation tests, whether involving space-perception or body-tilt judgments, subjects made considerable systematic errors which mostly betrayed tilt underestimation [Aubert effect (A effect)] and peaked near 130 degrees tilt. However, the A effect was much smaller in body-tilt estimates than in spatial pointing, implying that the underlying signal processing must have been different. Pointing results obtained with the visual and the oculomotor paradigm were not identical either, but these differences, which were task-related (horizontal vs. vertical), were subtle in comparison. The tilt-dependent pattern of random errors (noisy scatter) was almost identical in visual and oculomotor pointing results, showing a steep monotonic increase with tilt angle, but was again clearly different in the body-tilt estimates. These findings are discussed in the context of a conceptual model in an attempt to explain how the different patterns of systematic and random errors in external-space and self-tilt perception may come about. The scheme proposes that basically similar computational mechanisms, working with different settings, may be responsible.