Improving the robustness of Na¨ ıve Physics airflow mapping, using Bayesian reasoning on a multiple hypothesis tree. * Gideon Kowadlo Intelligent Robotics Research Centre Monash University Clayton, Victoria, Australia gkowadlo@ieee.org R.Andrew Russell Intelligent Robotics Research Centre Monash University Clayton, Victoria, Australia andy.russell@eng.monash.edu Abstract— Previous work on odour localisation in enclosed environments, relying on an airflow model, has faced significant limitations due to the fact that large differences between airflow topologies are predicted for only small variations in a physical map. This is due to uncertainties in the map and approximations in the modelling process. Furthermore, there are uncertainties regarding the flow direction through inlet/outlet ducts. We have presented a method for dealing with these uncertainties, by generating multiple airflow hypotheses. As the robot performs odour localisation, airflow in the environment is measured and used to adjust the confidences of the hypotheses using Bayesian inference. The best hypothesis is then selected, which allows the completion of the localisation task. We have shown experimentally that this method is capable of improving the robustness of our method for odour localisation in the presence of uncertainties, where previously it was incapable. The results further demonstrate the usefulness of Na¨ ıve Physicsfor practical robotics applications. Index Terms— Mapping, Na¨ ıve Physics, Odour Localisation, Bayesian, Multiple Hypothesis. I. I NTRODUCTION Odour localisation is performed by various organisms, and is now a burgeoning area of robotics research. Work in this field, and biorobotics in general, contributes to knowledge of biological organisms, and inspires innovation and new appli- cations. Robotic odour localisation experiments have explored and mimicked the chemotactic behaviour of such organisms as the silk worm moth (locate mates) [1], lobsters (find food) [2], dung beetles (find feces for brooding, living in or eating) and e-coli bacteria (find nutrients) [3]. Potential applications for odour localising robots include finding humans in search and rescue operations, locating gas leaks in industry, finding fires in their initial stages as well as many other possibilities. In general, odour particles are carried downwind, forming a plume that spreads, meanders and becomes patchy. Gradients of concentration and airflow are exploited using reactive behaviour by various organisms, and robots [3]–[5]. In some enclosed cluttered environments (low ceiling and thinly pop- ulated by objects that affect airflow) the plume is poorly defined, and odour dispersal is dominated by the formation of sectors. Odour can be confined to sectors (local or downwind * This is work is supported by the ARC funded Centre for Perceptive and Intelligent Machines in Complex Environments. from the release location). The reactive behaviour of previous solutions is not suited to, and therefore unreliable in these types of environments [6], which may be encountered in a cave, air duct, sewer or crawl-way beneath a house. In previous work, we have shown that the problem can be tackled effectively using a ‘sense-map-plan-act’ style control strategy [7], [8], comprised of firstly mapping the airflow in the environment using a Na¨ ıve Physics algorithm (NaReM) [9], and then using this map to reason about odour dispersal, plan an information gathering stage, navigate the environment taking readings, and finally make a prediction. Na¨ ıve Physics is the chosen method for airflow modelling as it avoids many of the difficulties of using computational fluid dynamics, and provides data structures at a high level of abstraction, readily used by the reasoning algorithm. In this project, the odour source location prediction was restricted to a physical domain. In order to make the localisation specific, vision was com- bined with olfaction to perform a bi-modal complementary sensing search [10]. For this approach to work effectively, the topology of sectors (from the airflow map), must be correct. This is very difficult to achieve, as small variations in the environment and in the implementation of the airflow modelling, can result in extremely different, but stable, airflow patterns. Furthermore, there are significant uncertainties in the information available to the robot prior to computing the airflow model, arising from the physical map provided to the robot, boundary conditions, and the application of the airflow modelling rules. Uncertainty leads to problems of data association, which have been tackled extensively in the areas of tracking, local- isation and mapping, with the use of Kalman filters, particle filters, and other methods based on Bayesian reasoning. Further, multiple hypothesis trees have been used since the late 60’s (see [11] for a good review), and combined with Bayesian reasoning by Reid [11]. Similar methods have been used for tracking [12], as well as for a variety of other applications from genetics (phylogeny) [13], to image processing (edge grouping and contour segmentation) [14], and mapping (modelling a dynamic environment) [15]. In this paper, we demonstrate an effective method for improving the robustness of the approach, in the presence of imprecise information and uncertainties. Where uncertainties