News & Events

Homotopy-Aware, Temporally Sampled Efficiently Adaptive State Lattices for Mobile Robot Navigation in Cluttered, Unrehearsed Environments

Ashwin Menon

Supervised by Thomas Howard

426 Computer Studies Building

An autonomous ground robot that operates in off-road environments needs to be robust to various challenges. Chief among these challenges is the environmental complexity that the robot is subject to when navigating to its goal. Due to this complexity and unpredictability of how the robot’s surroundings will look, it cannot rely on an a priori map to navigate in. Instead, it must use onboard sensing to build a sufficient representation, also known as a map, as it moves through the world to avoid hazards and reach its goal. 

This notion of partial observability is made even more challenging by recognizing that the map the robot builds is often imperfect and subject to incorrect or incomplete information. Robot sensors such as cameras, LiDARs, and depth sensors frequently build noisy world representations due to their own inherent sensing noise. If the robot is driving through a dense forest, many obstacles with the capability to immobilize the robot will not even show up on its map due to heavy occlusions from trees and logs. 

Classical robot autonomy architectures make use of these sensors in a system known as the perception module, which works in tandem with the state estimation module to accurately determine the robot’s position in the environment. With the robot’s position and a map of its surroundings, the regional planning module generates high level motion which is then supplied to a local planning module. The local planner, operating at a much higher frequency than the regional planner, will follow the regional planner’s reference trajectory towards the goal. This process repeats periodically until the robot reaches its goal.