Generalizing locomotion style to new animals with inverse optimal regression

We present a technique for analyzing a set of animal gaits to predict the gait of a new animal from its shape alone. This method works on a wide range of bipeds and quadrupeds, and adapts the motion style to the size and shape of the animal. We achieve this by combining inverse optimization with sparse data interpolation. Starting with a set of reference walking gaits extracted from sagittal plane video footage, we first use inverse optimization to learn physically motivated parameters describing the style of each of these gaits. Given a new animal, we estimate the parameters describing its gait with sparse data interpolation, then solve a forward optimization problem to synthesize the final gait. To improve the realism of the results, we introduce a novel algorithm called joint inverse optimization which learns coherent patterns in motion style from a database of example animal-gait pairs. We quantify the predictive performance of our model by comparing its synthesized gaits to ground truth walking motions for a range of different animals. We also apply our method to the prediction of gaits for dinosaurs and other extinct creatures.