SAUSAGES: Between Planning and Action

SAUSAGES stands for System for Autonomous Specification, Acquisition, Generation, and Execution of Schemata. SAUSAGES provides a framework for specifying, running, monitoring, and altering plans for mobile robots, and is ideal for building complex real-time systems that need to operate outdoors. SAUSAGES bridges the gap between planning and action, between the world of symbols and propositions and the world of pixels and actuators. On the one hand, system designers whose primary interest is perception can easily build a SAUSAGES system capable of carrying out complex missions that showcases their perception modalities and actuation modules without a lot of resources devoted to higher level task management and control. On the other hand, system designers who are primarily interested in planning can use SAUSAGES to provide a layer of abstraction for high level planners so that they do not have to worry about the implementation details of each plan step. This allows for the natural development of a robot that can act as an intelligent agent in the unstructured outdoor environment rather than just in simulation or in a highly structured laboratory environment. 1 Alternatives to SAUSAGES Several years ago, it became apparent that some form of mission execution and monitoring was needed to integrate the capabilities of the perception systems developed at Carnegie Mellon for the Navlab, a computer controlled Chevrolet vanll01. We had road followers that robustly followed roads[6], object detectors that avoided obstacles[91, landmark recognizers that localized the vehicle position in a map[3], but we had no consistent architectural glue to join them together. A robust road follower is impressive on its own, but a road follower alone has no way to know which way to turn at an intersection, no way to know when to speed up or slow down for important events, etc. A system that can execute a complex mission cannot simply be the sum of its perceptual modalities, there needs to be a “plan” which uses high level knowledge about goals and intentions to direct the behaviors of the low level perception and actuation modules. An intuitive way to represent a plan and execute it i s through events, e.g., “Follow this road until the second intersection, turn left, and stop at a mail box.” This plan representation works well for simulated and indoor mobile robots, which operate in a highly structured and controlled environment where every event necessary to the plan execution is easily perceived. Unfortunately, once a robot system leaves the laboratory it is in a much less stmctured, less controlled, and less predictable environment. The events that an outdoor mobile robot system needs to detect to execute a plan are very difficult to discriminate from the background environment. Reliable and accurate acquisition of perceptual cues, such as intersections or signs, is an almost impossible task. A system that relies solely on these perceptual events to direct its plan is not feasible. To get around this problem and still have a system that could execute interesting missions we developed Annotated Maps[ 111. An Annotated Map system reduces the problem of monitoring for arbitrary events, such as intersections or signs, to the problem of monitoring the vehicle position on an a priori map. The plan in an Annotated Map is essentially a positional production system, with productions like “at position x, perform action y” instead of productions like “if event x then perform action y.” This alternate representation of plans vastly reduces the computational resources necessary to perform a mission, while still allowing complex scenarios. We have extensively tested this system. For example, we used it to drive the Carnegie Mellon Navlab from one house in a suburb north of Pittsburgh to another approximately one kilometer away. The system used neural networks to drive the vehicle along the roads and used a laser range finder to detect obstacles and landmarks. The system successfully performed the mission, speeding up and slowing down at the appropriate places, navigating through several intersections, and coming to a stop at the proper mailbox. Figure 1: shows the annotated map we used in this system. We built this map by driving the vehicle over the course, storing the road position received from the controller and the landmarks detected by the laser range finder. A human expert then went through and added “annotations,” ortrigger lines. When the vehicle crosses these lines various actions are taken, such as setting speed, or turning modalities on or off, Thus, these trigger lines implicitly specify the plan. 2 Tech Report CMU-RI-TR-94-32 Figure 1: A sample annotated map The process of navigating through an intersection shows how this simple geometric implicit plan specification works. We had no perception system that allowed us to navigate an intersection, so it had to be done using only the map, and thus the vehicle’s position had to be accurate. Figure 2 gives an example of this process. As the vehicle’s position is tracked going down the. road, the first Uigger line that it crosses will turn on landmark based position updating to pinpoint the vehicle’s position on the map for precise navigation. Then the vehicle will cross a series of lines that will slow it down in stages. Once the vehicle actually enters an intersection, it crosses a line which tums off mad following and t u m s on dead reckoning from the map to actually make a left turn. Once the vehicle leaves the intersection it crosses lines that turn road following back on, turn landmark based position updating off, and sets the vehicle speed back to maximum again. Each step in this “plan” is sequenced implicitly by the positioning of the trigger lines and assumptions about the motion of the vehicle. Figure 2: Navigating an interseetion with an Annotated Map Resume road fdlowing Turn right using dead reckoning peed up to 4.0 m l s