If Abstraction Is the Answer, What Is the Question? — Reasoning for Everyday Manipulation Tasks

Abstract

In recent years we have seen tremendous advances in the mechatronic, sensing and computational infrastructure of robots, enabling them to act faster, stronger and more accurately than humans do. Yet, when it comes to accomplishing manipulation tasks in everyday settings, robots are still far away from reaching the sophistication and performance of humans. A key component of the “action intelligence” needed for reaching such a level of sophistication and performance are robot control systems that can take vague action descriptions and automatically infer how they are appropriately executed in a given task and environment context.

Artificial Intelligence (AI) is the research discipline that has studied such reasoning problems for more than fifty years. Researchers in AI have investigated naive physics reasoning, temporal projection, reasoning about action and change, action planning, spatial reasoning, to name only a few. Unfortunately, the proposed methods have not yet achieved their desired impact on autonomous robot control. We believe that one of the reasons is that most AI researchers consider perception and action to be the mere input and output of symbolic reasoning. In contrast, some researchers in cognitive psychology suggest a “simulation model” for reasoning through actions. In their view predicting the consequences of actions is very similar to executing the actions without causing physical effects — perception and action get simulated at a very fine-grained feedback loop.

In this talk I will present reasoning techniques of an autonomous robot control system that are inspired by the “simulation model” for reasoning through actions. These techniques use perception and motor control mechanisms and simulations thereof not only as input and output but more importantly also as resources for symbolic reasoning. I will show, using an autonomous robot making pancakes as an example, that such techniques reason about actions more realistically and thereby enable the robot to improve its performance.