Real-time Collision using AI
Each object is represented as a quadric decomposition – a collection of parts, where each one is an intersection of quadric inequalities. Collision detection, consisting of deepest point estimation and a prediction of an intersection polygon, is formulated as a semi-definite programming problem and solved using Recurrent neural networks. The method can be applied to rigid, elastic, articulated, or deformable bodies, modeled by both convex or non-convex quadrics. The model was trained on 100 million contact points with the expected relative prediction error 10-6 and 99.8th percentile below 10-5. The model does not need to be retrained for new scenes.