A neural network approach for real-time collision detection

The objective of the present work has been to develop a collision detection algorithm suitable for real-time applications. It is applicable to box-shaped objects and it is based on the relation between the colliding object positions and the impact point. The most known neural network (multilayer perceptron) trained with the familiar backpropagation learning algorithm has been used for this problem; such algorithm models the collision, then decides the impact point and the direction of the forces. The algorithm results are very good for the case of box-shaped objects. Furthermore, the computational cost is independent from the object positions and the way the surfaces are modeled, so it is also suitable for real-time applications. The model is being used and validated in a real harbor crane simulator developed by the Robotics Institute for Valencia Harbor in Spain.