Differentiable and modular prediction and planning for autonomous machines

This invention describes a smart system for self-driving machines that uses artificial intelligence to predict what other vehicles or pedestrians will do. Based on these predictions, the system figures out several possible paths for the self-driving machine, evaluates how good each path is, and then picks the best one. A key feature is that the entire process—from predicting others' movements to choosing a path and controlling the machine—is designed so that the AI can learn and optimize all these steps together, rather than treating them as separate problems. The claims specifically focus on a system that uses a neural network to predict future locations of agents, then calculates candidate paths and their costs, and finally computes a control sequence, all of which can be jointly trained.