Hello everybody, I am currently working on a system of 2 DOF robotic manipulators arranged in a crowded environment (i.e. the manipulators are very close to each others). Each manipulator has to reach a specific target position avoiding to hit the others manipulators during the movment. The strategy to obtain this behaviour is using a decentralized navigation function for each manipulator, basically each manipulator is attracted by its target and it is repulsed by the other manipulators (just like a potential field).
The problem is that with this strategy the manipulators can stumble in local minima and the convergence then is not reached. In the system are present hundreds of these manipulators and usually the percentage of manipulators that reach their target position is around 80-85 %.
My goal now is to develop a machine learning or deep learning algorithm that given a new position of the targets of each manipulator is able to predict the convergence of every single manipulator, on the basis of the past initial configuration of the targets and the following convergence.
Each manipulator is characterized by the parameters of its position in the plane, the parameters of its target and the label (0 or 1) that indicates if it has reached its target position or not.
I am not a data scientis but I am fascinated by this subject, and I would like to know what could be the best algorithm for this kind of problem.