Task planning contains finding a proper solution to a mission, from beginning to end that contains both discrete decisions (distribution of tasks according to entities, order of events) and continuous decisions (the time which each event is taking place at, planning the trajectory). The task planning problem is a complex problem with undecidable complexity. The ScottyActivity planner’s solutions contain the order of the events and the assignment of the continuous variables as well. These solutions do not consider the physical aspect due to the complex non-linear dynamics. Therefore, the solution provided is not necessarily the optimal one.
The purpose of this project is to build a tool based on unsupervised deep learning to plan the continuous problem when the order of the events is given, under physical constraints of speeds, accelerations and even time windows so that a feasible and physical trajectory is obtained. In the paper METHODS IN HYBRID PLANNING WITH DYNAMIC MODELS, which presents the described problem, the learned parameters are the speeds and times according to the constraints and the requirement for velocity continuity. In this project, the continuous learned parameters are the times and accelerations since it is possible to deduce the velocities and positions of the entities according to the physical equations and thus the requirement for continuity is met.
In this project, the architecture of the model is a chain of neuron networks. That way, the movement history of the entity and the various constraints are considered and the continuity in the movement is ensured so the solution would be physically correct. The loss function is defined in a way that weights all the different constraints defined in the task and gives penalties for exceeding them. As a result, the system manages to solve the optimization problem defined by the given mission and provides an optimal solution that meets all the requirements and constraints.