Althea: Recalled Process For Planning and Execution

We address two central problems causing a breakdown between planning and plan execution in robots. The first is the difficulty in reliably maintaining an accurate world model over time. Especially hard is the problem of perceptually segmenting objects and then tracking those objects. We suggest a representation language for plans that is based on perceptual data, not objects. The second major problem we address is the brittleness of plan operators. The execution of plan operators on actual robots has revealed the myriad of ways that the assumptions built into the construction of those operators can break. We suggest interleaving the execution of plans that contain sequences of motor and perceptual primitives, including operator failure. Having dramatically increased the size of each plan, we suggest a case-based representation as a means of keeping the search space tractable. based on vision, is not reliable. Even when segmentation is possible, the information generated is almost guaranteed to change without the planner’s knowing. For this reason, symbols in a classical plan that refer to objects in the world are often left dangling with no existing referent. In order to address the unreliability of perceptual segmentation of objects, we need planners that are embedded in autonomous agents to plan based on perceptual features, not objects. It is a tautology that a call to read the gripper’s width on a robot will yield the result that it does. It is an inference to conclude that there is an object in the gripper. This inference is not necessarily reliable. For instance a 66mm grip width may mean that there’s a Duplo in the gripper. Then again, the gripper might just be stuck open. A system that is limited to the predicate Gripping(Duplo-54) takes for granted that there