Machine-learning based system for path and/or motion planning and method of training the same

A method, comprising:

• generating a sequence of predicted occupancy grid maps (OGMs) for a set of trajectories and an expert trajectory for T-T1 time steps based on a sequence of OGMs for 0-T1 time steps, a reference map of an environment in which an autonomous vehicle is operating, the set of trajectories and the expert trajectory;
• generating by a neural network a cost volume for the sequence of predicted OGMs for each trajectory in the set of trajectories and the expert trajectory, each cost volume comprise a plurality of cost maps for T-T1 time steps, each cost map corresponding to a predicted OGM in the sequence of predicted OGMs and having the same dimensions as the corresponding predicted OGM, each cost map comprising a plurality of cells, each cell in the cost map representing a cost of the cell in corresponding predicted OGM being occupied in accordance with a policy defined by a policy function;
• determining an imitation cost for each trajectory in the set of trajectories by summing an imitation cost of each point for each time step;
• determining an imitation loss for the trajectory with a lowest imitation cost in accordance with a loss function which determines the imitation loss based on whether the state of the cells in the predicted OGM for the proposed trajectory differs from the state of the cells in the predicted OGM for the expert trajectory;
• determining an environmental cost for each trajectory in the set of trajectories by summing an environmental cost of each point for each time step;
• determining an environmental loss for the trajectory with a lowest environmental cost in accordance with a loss function which determines the imitation loss based on whether the state of the cells in the predicted OGM for the proposed trajectory differs from the reference map;
• determining whether a combination of the loss functions for the imitation loss and the environmental loss are below a threshold;
• and
• updating parameters of the neural network in response to a determination that the imitation loss and environmental loss is not below a threshold.