Gravity acts vertically downwards in the real world. By internalising that reality as an internal model we should perform better when relating to objects ‘falling’ naturally downwards rather than in some other direction. Several studies have shown that expectation of gravity’s effect on vertically descending or ascending targets can over-rule visual information to the contrary (1-3), but this has not been explored for other ‘unnatural’ directions. We sought evidence for an internal model of gravity by comparing visually guided tracking of objects accelerating at g (9.81m/s2) in straight lines vertically downwards, upwards and horizontally. If a model exists, differences should exist between tracking natural versus unnatural trajectories. For each direction, 17 subjects executed 3 tracks of a 15cm diameter realistic visual target ‘falling’ from stationary for 2.475m. The target was back-projected onto a screen 3m ahead of subjects. Tracking was by extending the dominant hand’s index finger and pointing at the moving target. Tracking intercept (indicated position on the screen) was determined using a vector from the dominant eye past the finger-tip to the screen. Vicon tracked movements. The instantaneous error (tracking position minus target position) was summed in 10ms steps over the time taken for the target to traverse 2.475m (730ms) as a measure of performance. Tracking ahead produced negative error values, and tracking behind positive values. Direction significantly affected error (one-way ANOVA: F=31.57, p<.001). Post hoc tests revealed all errors were significantly different from one another (p<.001): descending M=-8.405m SD=2.342; ascending M=18.426m, SD=10.022, horizontal: M=7.755m SD=6.965). Tracking, the output of the manual tracking controller, was below the target: leading descending targets and lagging behind ascending targets. Starting errors were also below the target in all conditions. To judge one’s tracking position, one is required to keep both finger-tip and target in sight. Due to the physical presence of the arm/hand/finger one must ensure that the target is above the tracked location to prevent total or partial target occlusion. Error when tracking a target accelerating upwards was more than double that for a target accelerating naturally downwards, and tracking fell behind for a horizontally accelerating target. These different errors for different directions suggest that subjects expected non-descending (rising or horizontally moving) targets would not accelerate in the same way as a free-falling target. This pattern of results suggests subjects used an internal model of gravity to predict the future position of ‘free-falling’ targets and that the model is directionally tuned. Also, such tuning is evidence that subjects did not use either feedback or direct perception when tracking.