Off-Road Obstacle Avoidance through End-to-End Learning