Robot Navigation in a Known Environment with Unknown Moving Obstacles

To be useful in the real world, robots need to move safely in unstructured environments and achieve their given goals despite unexpected changes in their surroundings. The environments of real robots are rarely predictable or perfectly known so it does not make sense to make precise plans before moving. The robot navigation problem can be decomposed into the two problems of getting to a goal and avoiding obstacles. The problem of getting to a goal is a global problem in that short paths to the goal generally cannot be found using only local information. The topology of the workspace is important in finding good routes to a goal. The problem of avoiding obstacles can often be solved using only local information, but for an unpredictable environment it cannot be solved in advance because the robot needs to sense the obstacles before it can be expected to avoid them. Some have solved the navigation problem by solving these two sub-problems one after the other. A path is first found from the robot’s initial position to the goal and then the robot approximates this path as it avoids obstacles. This method is rather restrictive in that the robot is required to stay fairly close to or perhaps on a given path. This would not work well if the path found goes through a passageway which turns out to be blocked by some unforeseen obstacle. Solutions that are only local, such as those produced often by artificial potential fields, often lead the robot into local minima traps. We propose a much more flexible solution using a common tool, the artificial potential field, in a new form that we call a "hybrid artificial potential field". A hybrid potential field is obtained by combining two different kinds of artificial potential fields, a global discontinuous potential field and a local continuous potential field. The global potential field covers the whole floorplan and captures the static floorplan information. Since it includes only information about static objects it can be computed a priori when given the goal. We represent this field as a two dimensional array of heights so that the robot rolls downhill to the goal while avoiding fixed obstacles. In this potential field all free cells will be assigned the city block distance of the shortest free path from that cell to the goal so this field has no minima other than at the goal. The local field captures local information obtained by sensors about obstacles and covers only the area around the robot. The purpose of this field is to push the robot away from obstacles that are on its path. This field has to be computed repeatedly as the robot moves in the workspace. We continuously sense the environment with sonar sensors and regularly update the dynamic potential field. For each sensor reading we place a "hill" in the potential field. Since these hills are built as the obstacles are sensed this method works with obstacles that move in unknown, unpredictable paths. Obviously, if obstacles move too fast collisions are sometimes unavoidable. Unfortunately, when these two types of potential fields are added together, the result may have local minima. We have analyzed the causes of these minima and found ways to deal with them.