Michael Beetz

Institute for Artificial Intelligence
Bremen University, Germany

Title: Robot planning for the mastery of human-scale everyday manipulation tasks

Abstract: Robot planning can be considered as the reasoning about the future execution of robot programs (plans) in order to optimize their performance in terms of achieving their goals and efficiency (McDermott). The holy grail of robot (action) planning ever since the Shakey project has been to equip robotic agents with human-level (manipulation) action capabilities. Unfortunately, the progress along this dimension has been modest at best. I believe that much of the lack of progress is caused by the way the research eld of task planning abstracts (robot) actions (see PDDL). It makes the assumption that reasoning about abstract preconditions and e ffects is sufficient for planning complex manipulation tasks. If we interpret this assumption from a probabilistic point of view, we can restate it by asserting that the probability of achieving the desired e ffects of actions is conditionally independent of how the robot executes the actions given that the preconditions of the actions are satis ed. This means that our robot action planning systems would not change their belief about whether an action is executed successfully depending on whether the robot plans to grasp an object with one hand or two, which grasp type it applies, and so on. Or, if a fetch action is executed by two-year old or an experienced waiter. In contrast our experience with realizing human-scale manipulation activities for robotic agents shows that most of the intelligent problem-solving capabilities of robots are needed in order to decide how to execute the actions to make them succeed, that is to achieve the desired e ffects of an action and avoid the undesired ones.
I believe that in order to materialize the impact that robot planning technology can have for robotic agents that are to accomplish human-scale manipulation activities, we have to extend our representation and reasoning machanisms to include the concepts of motor cognition. Motor cognition is a discipline in cognitive psychology of action which is concerned with the learning, reasoning, and planning of how to parameterize and synchronize motions in order to accomplish actions. I foresee a new generation of powerful robot planning systems that do not only reason symbolically about their actions but also subsymbolically with their \eyes and hands”. Today’s disruptive technologies, in particular modern game technology, physics simulation, data analytics, and deep learning give us the opportunities to pursue this direction.

This entry was posted in Uncategorized. Bookmark the permalink.