This module introduces the richness and challenges of the self-driving motion planning problem, demonstrating a working example that will be built toward throughout this course. The focus will be on defining the primary scenarios encountered in driving, types of loss functions and constraints that affect planning, as well as a common decomposition of the planning problem into behaviour and trajectory planning subproblems. This module introduces a generic, hierarchical motion planning optimization formulation that is further expanded and implemented throughout the subsequent modules.

The occupancy grid is a discretization of space into fixed-sized cells, each of which contains a probability that it is occupied. It is a basic data structure used throughout robotics and an alternative to storing full point clouds. This module introduces the occupancy grid and reviews the space and computation requirements of the data structure. In many cases, a 2D occupancy grid is sufficient; learners will examine ways to efficiently compress and filter 3D LIDAR scans to form 2D maps.

This module develops the concepts of shortest path search on graphs in order to find a sequence of road segments in a driving map that will navigate a vehicle from a current location to a destination. The modules covers the definition of a roadmap graph with road segments, intersections and travel times, and presents Dijkstra’s and A* search for identification of the shortest path across the road network.

This module introduces dynamic obstacles into the behaviour planning problem, and presents learners with the tools to assess the time to collision of vehicles and pedestrians in the environment.

This module develops a basic rule-based behaviour planning system, which performs high level decision making of driving behaviours such as lane changes, passing of parked cars and progress through intersections. The module defines a consistent set of rules that are evaluated to select preferred vehicle behaviours that restrict the set of possible paths and speed profiles to be explored in lower level planning.

A reactive planner takes local information available within a sensor footprint and a global objective defined in a map coordinate frame to identify a locally feasible path to follow that is collision free and makes progress to a goal. In this module, learners will develop a trajectory rollout and dynamic window planner, which enables path finding in arbitrary static 2D environments. The limits of the approach for true self-driving will also be discussed.

Parameterized curves are widely used to define paths through the environment for self-driving. This module introduces continuous curve path optimization as a two point boundary value problem which minimized deviation from a desired path while satisfying curvature constraints.