Learning obstacle avoidance reflex behavior for autonomous navigation from hand-drawn trajectories

The present work explores a simple off-line method to extract the intuitive actions used by humans to avoid obstacles during motion in unknown environments. The proposed method analyzes the hand drawn trajectories by human individuals on environment maps showing typical obstacle placements, and evaluates the navigational decision parameters. The translation and steering velocity variation along the curve are computed based on the constraints of the mobile entity (e.g., an autonomous mobile robot). The decisions are considered to be taken in the context of the distances of the obstacles around the current point on the trajectory. The instances of environmental situations and corresponding intended actions are used to train a neural network. To reduce the complexity of the network, the number of input variables for the network is reduced by considering only single sided reflex behaviors. The left-right symmetry of the perception-action behaviors allows the single sided reflex network to be used for both left and right hand side reflex in the vicinity of obstacles. Simulation results are presented to show the effectiveness of the proposed strategy in typical obstacle situations.