The global dynamic window approach is a powerful motion primitive for mobile robots. It does not address the problem of mapping and localization, but it provides an extremely powerful method to move a mobile robot to a goal location in a dynamic and cluttered environment. We performed experiments on on a holonomic mobile base (Nomadic XR4000) and achieved very robust and fast motion generation.

By combining a very fast global path planning method with local execution schemes, reactive behavior is achieved without sacrificing goal-directed motion. The local execution scheme is based on the dynamic window approach by Fox, Burgard, Thrun. It determining a motion command the method takes the dynamic properties of the robot into account. Only those motion commands are executed that will allow the robot to come to a stop before hitting an obstacle. Most global path planning methods only provide a direction of motion and not a velocity. Using the dynamic window approach, we can choose an adequate velocity in the direction of motion given by the path planner.

The global path planner is based on the navigation function NF1 by Barraquand and Latombe. It builds a local minima-free potential function in an occupancy grid (see figure: navigation function in red, robot in blue, goal in green) using a wave-propagation approach. The integration of global path planning and execution methods computes the navigation function during every servo tick. The navigation function is not computed for the entire world, but only for those regions required (see red gradient in the figure). This helps to achieve real-time performance, even in large and complex environments. Each time the local execution method chooses a motion command, those commands moving along the gradient of the navigation function are preferred. This achieves a very simple, yet powerful integration of global and local methods.