On the Role of Quantitative Descriptions of Behaviour in Mobile Robotics Research

This paper — a summary of a keynote address given at the Robocup 2003 symposium — argues i) that mobile robotics research would benefit from a theoretical understanding of robot-environment interaction, ii) that independent replication and verification of experimental results should become common practice within robotics research, and iii) that quantitative measures of robot behaviour are needed to achieve this. The paper gives one example of such quantitative measures of behaviour: the reconstruction of the phase space describing a robot’s behaviour, and its subsequent analysis using chaos theory. 1 Mobile Robotics Research 1.1 Robot Engineering versus Robot Science Theory Supports Design Arguably, there are (at least) two independent objectives of robotics research: on the one hand, to create artefacts that are capable of carrying out useful tasks in the real world — for example industrial, service, transportation or medical robots, to name but a few, and on the other hand to obtain a theoretical understanding of the design issues involved in making those artefacts — for example sensor and actuator modelling, system identification (modelling of entire systems), or sensor, actuator and behaviour analysis. The former can be referred to as ‘robot engineering’, the latter as ‘robot science’. It is robot science that this paper is concerned with. While robot engineering ultimately produces the ‘useful’ artefacts, there is a lot that robot science can contribute to this process. Without theoretical understanding, any design process is largely dependent upon trial-and-error experimentation and iterative refinement. In order to design in a principled way, a hypothesis of some kind (a justified expectation) is needed to guide the design process. The hypothesis guides the investigation: results obtained are fed back into the process and brought into alignment with the theory, to lead to the next stage of the experimentation and design. The better the theory underlying the design process, the more effective and goal-oriented the design process will be. Every process of designing technical artefacts is based on some kind of assumptions (a ‘theory’), even if very little is known at all about the object being 0 Published in Proc. RoboCup Symposium, Padua 2003, Springer Verlag. designed. This is true for current mobile robotics research, too. When asked to design a wall-following robot, the designer will not start with an arbitrary program, but with a ‘reasonable guess’, sensibly speculating on which sensors might be useful to achieve the desired behaviour, which general kind of control program will perform acceptably, etc. But, given our current understanding of robotics, he is unable to design the entire behaviour off-line! Instead, mobile robotics researchers to-date are crucially dependent on trialand-error procedures. A ‘reasonable prototype’ has to be tested in the target environment, and refined based on observations and underlying theory (‘hunch’ is often the more appropriate term for such theories). Here is a practical example: to design the Roomba commercial robot floor cleaner (relying on very simple sensing, and not involving any sophisticated navigation), thirty prototypes had to be built over a period of twelve years [EXN 03]! The first argument we would make in favour of a better theoretical understanding of the principles underlying a mobile robot’s operation in its environment, therefore, is that robot engineering (the process of designing a technical artefact that will perform useful tasks in the real world) will benefit from theory through the resulting more effective, rigorous and goal-oriented development methods. Science Requires Replication and Verification Current mobile robotics research practice not only differs from that of established disciplines in its lack of theories supporting design, but also in a second aspect: independent replication and verification of experimental results is uncommon. While in sciences such as biology or physics, for instance, reported results are only taken seriously once they have been verified independently a number of times, in robotics this is not the case. Instead, papers often describe experimental results obtained in specific environment, under specific experimental conditions. These experiments therefore are ‘existence proofs’ — the demonstration that a particular result can be achieved — but they do not state in general terms under which conditions a particular result can be obtained, nor which principles underlie the result. Existence proofs are useful, they demonstrate that something can be achieved, which is an important aspect of science, but they do not lead towards general principles and theories. The second argument we make, therefore, is that mobile robotics research is now at a stage where we should move on from existence proofs to a research culture that habitually includes independent replication and verification of experiments. The Role of Quantitative Descriptions Theories, experimental replication and experimental verification all depend crucially on quantitative descriptions: quantitative descriptions are an essential element of the language of science. The third argument we make, therefore, is that an essential first step towards a sounder theoretical understanding of robot-environment interaction is to develop and apply quantitative descriptions of robot-environment interaction. The experiments reported in this paper are one example of how to achieve this.