Gradient Descent Dynamic Window Approach to Mobile Robot Autonomous Navigation

Abstract

Avoiding obstacles is a key feature in a vehicle autonomous navigation methodology. The Dynamic Window Approach (DWA), proposed several decades ago, has emerged as a responsive navigation methodology suitable for reactive obstacle avoidance. In the initial approach of the DWA, the optimization of an objective function is realized with an exhaustive computation, which can be costly in computational time. This is not useful in a real-time scenario where an autonomous vehicle needs to avoid obstacles in urban or road velocity conditions. In this paper, we revise the DWA methodology by implementing a new method that redefines the objective function as a loss function, allowing the application of gradient descent for optimization. We verified the correctness of our optimization by controlling a robot in an unknown environment with obstacles to visit given positions. Our approach was tested in simulation on ROS and on a real Turtlebot robot, demonstrating improved runtime execution and a less abrupt, more comfortable driving experience.