Motion planning is one of the well researched fields in the robotics domain. In recent years, researchers have been able to solve complex planning problems which were considered unsolvable or too hard to solve for a long time. For an intelligent robot to accomplish a complex task, sometimes it needs to plan through spaces which consist of intersecting submanifolds and also have different dimensionalities. Consider a use case, in which a robot needs to pick a plate from a table. Due to its physical constraint, it can only pick the plate by pushing it to the end of the table. Traditionally, the entire task is divided into subtasks: moving towards table, pushing the plate towards the end of the table and picking the plate. Then, those subtasks are solved independently. As those subtasks are manually defined, such a divide and solve method is not scalable and cannot be used in a different environment. In order to solve such a task, the planner needs to plan through different discrete actions in order to obtain a continuous motion. This thesis focuses on such a specific planning problem called "multi modal planning problem". The proposed work is divided into two main parts: a multi-modal planner and a trajectory prediction. In order to solve the multi-modal planning problem, a global planner based on a sampling-based algorithm is used to plan through diverse discrete actions. Even though such a sampling based planner provides a feasible global path, an optimal path is not guaranteed. To achieve a smooth optimum path and to replan its path in case of a dynamic obstacle or uncertainty from the sensor reading, a path integral based stochastic optimizer is used as a local planner. In combining global and local planner, the obtained path will be a globally optimal path, which has the ability to replan its path if necessary. In the second part of the work, a trajectory prediction method is proposed for the multi-modal planner. It is a common procedure to start the planning process from scratch for every new task without exploiting similarities to previous situations. The trajectory prediction method provides an initial trajectory for the planner which was learnt from the trajectory used in similar situations. By predicting such an initial solution the entire multi-modal planning process will be speed up.