Knowledge Representation and Reasoning
Michael Beetz · Cognitive Robotics · 2022
Robots are already making large strides in their abilities, but as the generalizable knowledge repre senta tion prob lem is addressed, the growth of robot capabilities will begin in earnest, and it will likely be explosive.The effects on economic output and human workers are certain to be profound.-Pratt 2015 IntroductionOne of the most impressive cognitive capabilities of humans is the ability to accomplish their everyday manipulation tasks.In most cases, simple and vague instructions such as "set the table," "bring me something to drink," or "clean up" suffice to let us know what to do.The be hav ior that humans generate in order to perform such manipulation tasks is sophisticated, complex, and tailored to the objects they manipulate, their skill level, the context of the task, and the surrounding scene in which the task is to be performed.Accomplishing these tasks also requires humans to avoid common pitfalls such as breaking objects or spilling fluids.A main challenge in accomplishing a task such as "set the table" is that it is underdetermined.The request does not spell out which objects to put on the table, the arrangement of the objects, where to find the objects, how they look, how they have to be handled, how they can be efficiently carried, or whether there are social conventions on how to grasp and hold them.Consequently, humans must have the knowledge and the reasoning capacity required to close the gaps between what they are explic itly told and what they are expected to do.This knowledge, including commonsense and intuitive physics knowledge, is shared by most humans, which makes it pos si ble for a person to execute a task to the satisfaction of the person requesting it even if the instructions are vague.By contrast, imagine how hard it must be to write a robot control program for an autonomous house hold assistant robot that has to accomplish these tasks in dif fer ent house holds, with dif fer ent objects, for dif fer ent habits and preferences, and under dif fer ent circumstances, requiring the program to select the most adequate course of action in so many pos si ble contexts. 1 Many dif fer ent approaches can be taken to generate robot control programs for tasks such as "set the table," including robot learning (Peters et al. 2016), task and motion planning