Project Overview

Motivation

Path planning is a well-known problem in the field of robotics. Autonomous vehicles constantly face the issue of navigating through novel scenarios, and while roads remain fairly consistent, obstacles that come up change constantly. Other planning problems can take place in new environments. In search and rescue scenarios, different regions might end up more navigable than others, and a robot would have to determine the best course of action to avoid obstacles, conquer rough terrain, and reach its destination.

Navigating the real world often takes a lot of guesswork. In mission critical scenarios, minimizing the uncertainty of a chosen path is crucial to safety. A poor navigational decision can cripple a robot, so it must avoid areas that might be seen as more dangerous. While no system is born with knowledge about safety and drivability, making such an identification and learning process autonomous can improve performance during eventual deployment.

The Pitch

Our goal is to create a system that performs this kind of terrain-aware navigation. Using a Turtlebot and its camera, we create a map of the environment as we navigate. Simultaneously, a feature vector is updated with information about the regions the robot has taken an image of so as to inform the control algorithm about the terrains present. As the robot physically drives over different areas, data about aleatoric (environmental) and epistemic (model-based) uncertainties are recorded to inform the path planner. A voxel grid is constantly updated with edge weights corresponding to these uncertainties, and the planner can find optimal paths through it. Consequently, we can go through regions with different terrains while choosing the most navigable route.