EPSRC grant reference EP/D054419/1
Principal Investigator - Dr Myra S. Wilson
Co-investigator - Dr Joanne Walker
Tel: +44 1970 622928
Fax: +44 1970 628536
Email: mxw@aber.ac.uk
The need for autonomy in the area of mobile robotics is increasing. With more applications relating to areas where robots may be out of contact for extended periods of time (such as during planetary or geographically remote earth-based exploration tasks), it is becoming important that they continue to work effectively and autonomously adapt to different environments without performing inappropriate actions which could jeopardise the robot or its surroundings. Most current systems are grounded in one environment and their competence breaks down when either the environment changes, or on robot degradation. What is required is a robot system that will need little or no on-site programming and which can adapt within limits to its own degradation and to changes in its task or environment to keep going as long as possible without failure. The robots must also deal intelligently with any failures which do occur. Other criteria, such as ensuring that power constraints are met, should also be addressed.
A multi-robot group consisting of heterogeneous robots offers several advantages over single robots or homogeneous teams. These include the ability to use their individual capabilities to deal with a more diverse set of tasks, to complete a task when an individual robot fails through its own degradation, or to take over a task if environmental changes mean that the original robot assigned to the task can no longer perform it.
This proposal envisions a multiple, heterogeneous robot system that is able to operate without external assistance for an extended period of time. This highlights the importance of adaptability, usefulness in diverse environments, and the advantage of not having to perform complex reprogramming. The system should be able to identify when a robot has degraded and is unable to complete a task. The task can then be reallocated to another robot with a similar competence. Where the environment has changed and the robot is no longer appropriate for the task, this should be recognised and the task reassigned to another robot and the original robot assigned elsewhere. This means that the system must have redundancy and adaptability during its lifetime, must be able to dynamically reallocate tasks and recognise task failure, and must have a large degree of autonomy. The proposal will not deal with multiple robots concurrently cooperating to perform a single task as this is a complex issue in its own right. An individual robot will attempt a task and will move away on failure or be treated by the other robots as a new obstacle to be dealt with.
The project intends using genetic and evolutionary computation (GEC) methods to achieve the above adaptation and failure recognition. GEC algorithms are a biologically inspired method of optimisation and search based on the concept of natural selection. Although GECs have been applied to the design of robots and their controllers in the past, they have rarely been applied to the problem of lifelong optimisation and adaptation of robots in dynamic environments. Evolutionary methods differ from most robot learning techniques in two main ways: firstly the amount of supervision required is usually lower, and secondly, the evolution does not impose constraints on what can be evolved. Genetic algorithms and evolutionary strategies have already been shown to enhance the robustness of the behaviour of a robot in changing environments.
Based on Walker's PhD thesis, the principal aim of this project is to develop a heterogeneous multi-robot system which is capable of a large degree of adaptation in diverse environments while performing useful task reallocation on catastrophic robot degradation or environmental/task changes. To do this, the project funds, firstly, a PhD student who will research lifelong adaptation in diverse environments for individual robots. Secondly, a PDRA who will concentrate on the multiple, heterogeneous robot system for recognising task failure or inappropriate behaviour, and robot reassignment, obtaining information through an evolutionary training phase.