A key challenge for the next generation of autonomous robots is the reliable and efficient accomplishment of prolonged, complex, and dynamically changing tasks in the real world. One of the most promising approaches to realizing these capabilities is the plan-based approach to robot control. In the plan-based approach, robots produce control actions by generating, maintaining, and executing plans that are tailored for the robots' respective tasks. Plans are robot control programs that a robot can not only execute but also reason about and manipulate. Thus a plan-based controller is able to manage and adapt the robot's intended course of action – the plan – while executing it and can thereby better achieve complex and changing goals. The use of plans enables these robots to flexibly interleave complex and interacting tasks, exploit opportunities, quickly plan their courses of action, and, if necessary, revise their intended activities. One of the grand visions in the area of plan-based robot control is the realization of general autonomous robot control programs that can adapt themselves to the environments they are to operate in and to the distribution of complex tasks they are to perform. An instance of this grand vision is a pre-programmed household robot that knows how to clean a kitchen, how to operate a dishwasher, and so on. Being installed in a new environment it specializes its general plans to the specifics of the household and learns to manage the specific agenda of household chorus that is given to it. The robot also has to learn about the pitfalls of its tasks and its environment and avoid them through foresight.

We investigate the plan-based control of autonomous mobile robots performing everyday pick-and-place tasks in human environments. Our approach applies AI planning techniques to transform default plans that can be inferred from instructions into flexible, high-performance robot plans. To find high performance plans the planning system applies transformations such as carrying plates to the table by stacking them or leaving cabinet doors open while setting the table, which require substantial changes of the control structure of the intended activities.

Cognition requires technical systems to reason about and revise their own control programs. In particular, the systems must be capable of predicting the effects of their intended courses of action, learning routine controllers from experience, including advice into the behaviour specifications, and explaining their own behaviour. Having such capabilities for complex and changing activities requires technical systems to form, maintain, and execute plans - control programs that cannot only be executed but also reasoned about, generated, and revised during their execution. Indeed, it is almost impossible to imagine that cognitive technical systems performing non-trivial, dynamically changing, and possibly interfering tasks could be successful without performing plan-based control.

The Cogito project builds on Structured Reactive Controllers (SRCs), one of the leading-edge plan-based robot control systems for autonomous service robots developed by our research group. Cogito develops the next generation of plan-based controllers that differs from the current generation in that it provides built-in mechanisms for all cognitive capabilities listed above.

Key contributions to the research area of plan-based control of robotic agents is the application of transformational planning and learning to concurrent reactive manipulation tasks. In this approach transformational planning cannot only be used to find plans faster but also to improve the flexibility, reliability, and performance of plans.

We have shown in experiments that using the Cogito approach we can specify reliable robot plans for everyday activity that are general, flexible, and efficient. Indeed we have shown that our plans can recover from 86% of local failures with the remaining failures being ones such as an object slipping out of the gripper and falling to an unreachable position. Improving the plans by changing the problem-solving strategy requires the planner to reason through plans that are on average several thousands lines of code. We have demonstrated that in spite of this complexity transformational planning and learning can improve the plan performance by 23% to 43% depending on the tasks using very general plan transformation rules.

This project was partly funded by CoTeSys.