Hierarchical Memory Organization of Multimodal Robot Skills for Plan-based Robot Control

Abstract

Today, autonomous service robots are still far from being part of our everyday life. This is hard to understand if one considers the remarkable success of robots in automation industries. In controlled environments the robots’ speed, accuracy and reliability by far exceed human capabilities. So the question arises: Why can robots not perform simple tasks in natural human environments? One reason for this inability is the lack of versatility of robot behaviors. Many engineers focus on increasing the performance of specific robot tasks and build robot hardware for certain applications to perceive and manipulate the environment more and more skillfully. In contrast, only few researchers investigate the versatility of robot hardware and software for different tasks. Another reason for the limited success of service robots is their lack of cognitive competence and flexibility. The ability to reason about tasks and problems is a key property for the application of robots in environments that cannot be preconceived completely by the developers. However, artificial intelligence (AI) techniques, that have remarkable success in several areas, are rarely applied to embodied robots, which may be down to the different representational formats of robotics and AI. Embodied robots inherently deal with continuous data while AI traditionally employs symbolic world representations. The core work of this thesis concerns the integration of different results from robotics and AI research into one coherent system that marks the state of the art of autonomous intelligent service robots. Integration takes place along two dimensions: First, several robot sensor and actuator modalities are integrated at robot skill level. While most sensors are developed with particular regard to certain applications and thus show different properties in the workspace, accuracy and measured physical modality, they can complement one another and provide a useful coherent picture of the robot’s environment if combined properly. As a result of this thesis, it will be shown that robot perception can particularly gain from multi-sensor integration if sensors are actively focused on objects of interest. Second, representation and planning techniques from the area of AI will be integrated with the developed multi-modal robot skills to establish a complete robot system from low-level perception and action to abstract reasoning and planning. Specifically, basic robot skills will be organized hierarchically into more abstract robot tasks which can be used on the planning level to decompose complex instructions into executable parts. Basic robot skills constitute the interface between planning and execution layer. The abstraction level of these skills determines the complexity of the two layers. Although every layered robot architecture has to deal with the question of which task should be achieved at the skill level and which task requires deliberative planning, it has not been addressed explicitly in the literature. In this thesis, properties of basic robot skills are discussed in detail, which leads to a number of arguments that constitute a guideline for the design of such basic skills in the context of layered robot architectures. The integration of multiple sensor and actuator modalities with state of the art AI planning techniques provides the necessary means for reliable robot behavior in complex situations. It will be shown that the proposed architecture leads to extended autonomy and makes abstract robot tasks possible. The TAMS service robot is used to demonstrate the system integration and to validate the results of this thesis. It consists of a mobile platform, a robot arm equipped with a three-finger hand and a rich sensor system. This setup enables the robot to perceive its environment and to perform everyday tasks autonomously. Endowing the robot with planning capabilities and integrating results from previous investigations is part of this work. Be