Abstract
People can perceive and predict physical properties of objects, even in complex and unfamiliar situations; for instance, judging whether a stack of several objects is stable or unstable is usually not difficult. These judgments suggest the brain has an internal understanding of physics concepts such as gravity, friction, and inertia. Some have argued that humans' “intuitive physics” is biased and poor when interpreting inertial dynamics, but the breadth and sophistication of humans' general visual physical reasoning is at odds with these accounts. Our work measures the boundaries of humans' stability perception to uncover the underlying cognitive mechanics that facilitate these behaviors.
We conducted a set of human psychophysical experiments in which participants viewed virtual 3D scenes of different configurations (“towers”) of ten blocks. Participants judged whether each tower was stable or not, which direction/how far it would fall if it was unstable, and what external force magnitude would be required to cause it to collapse if it was stable. We compared participants' judgments to a model “physics perception” observer that made these judgments by simulating true physical dynamics with minor positional noise added to the tower's individual block locations, and found people and model were highly consistent. We tested for learning by providing visual feedback of whether the tower collapsed to one group of participants, and no feedback to another, but found no significant differences between the groups. Several possible heuristics exist for performing this task (e.g. tower height proportional to instability), but we ruled them out because they underperformed the model predictions, and because participants made model-consistent judgments in a same-height tower control task. These results suggest humans apply concepts like gravity, solidity, and support in an approximate dynamics simulation to make physics judgments, rather than relying on weak, biased, special-case heuristics.