Bayesian theories of visual perception postulate that the brain infers from retinal images the accurate spatiotemporal properties of physical events by combining noisy sensory information with prior knowledge of environmental properties, such as Newtonian laws of motion. We tested this general assumption with a study where participants observed the trajectory of a ball rolling off a surface. We simulated this dynamic event in virtual reality: participants viewed a marble-sized ball fall towards them from a tabletop. The ball disappeared upon hitting the floor. Their task was to simply indicate the remembered location where the ball fell by positioning a probe dot on the ground floor below the tabletop. There were two sources of visual information specifying the final location of the ball. First, the final location itself, which was perfectly visible to the participants, although for a very brief period. Second, the parabolic motion of the ball, which provided predictive information about the possible fall locations. Moreover, prior information about Earth’s gravity could also have potentially improved the accuracy in the estimate of the ball’s trajectory. Contrary to the predictions of Bayesian models, we found that instead of improving perceptual accuracy, adding sensory information of the ball’s falling trajectory led to systematic biases in the judgement of its final location. Prior knowledge of Earth’s gravity clearly influenced the perceptual judgements, but did not lead to improved accuracy. We explain these results with a new theory of cue integration termed Intrinsic Constrained that does not postulate accurate perception of physical events.