Parameter Selection in the Dynamic Window Approach Robot Collision Avoidance Algorithm using Bayesian Optimization

The Dynamic Window Approach algorithm is a classical local collision avoidance algorithm used in mobile robot navigation to generate obstacle free trajectories that are feasible based on the motion dynamics of the robot. Trajectory selection is guided by a navigation function, a sum of sub-functions that evaluate the Heading direction, Distance to obstacle and Forward velocity of the robot. These sub-functions are each weighted using parameter constants within the interval {0,1}. Parameter weights play a role in determining the overall success of navigation in terms of reaching the goal point. In this thesis, Bayesian optimization, a Machine - Learning based optimization method for expensive black box functions, is applied to select optimal parameters that yield goal reaching trajectories within a shorter time span. A navigation simulation was built in MATLAB to model the motion dynamics of a rigid point in a 2-D Cartesian space coupled with obstacle avoidance based on the Dynamic Window Approach algorithm. A surrogate model was developed based on Gaussian Process Regression using a training set of 20 initial evaluations comprising of metrics that define the navigation simulation. Furthermore, an expected improvement function is iteratively sampled over the surrogate model to yield parameter sets that minimize a cost function f∗. Results show that the optimal parameter weights generated from the optimization process all yielded successful goal reaching navigation outcomes in simulation environments with varying number and arrangements of static obstacles. Also, minimizing navigation time penalizes the Distance to obstacle cost, as a result, resulting trajectories tend to be close to obstacles within the navigation space.