Computational Neuroscience & Machine Learning

A novel robotic system could be used to persistently inspect and monitor buried pipe networks, more accurately and precisely mapping and locating faults in the network, allowing efficient maintenance and providing valuable asset data.

This research is on localization for the robotic system, the problem of how a robot estimates its trajectory in the environment given uncertain position measurements. Probabilistic robot localization, typically based on Bayesian inference, has been well studied for open environments using visual, inertial, global positioning, and rangefinder sensing. However, the size, inaccessibility, and homogeneity of the buried pipe environment limit the effectiveness of these conventional methods. The environment also limits size and power for computation and communication between robots. The aim is therefore to develop a novel localization method, substantially different from existing methods, which is designed to solve the non-standard localization problem with the constraints described here. While this method will be applicable to all types of buried pipe network, the challenges described are more difficult in the case of water distribution networks, so this work is expected to be more impactful in this case.

The localization method will be derived from the same probabilistic principles as typical methods, addressing the above constraints:

(i) The localization will be done in the semi-discrete space of the water pipe network, rather than the typical continuous space of robots, using an appropriately defined coordinate system, allowing better modelling of position.

(ii) Systematic uncertainties can be modelled in the confined semi-discrete space without the increase in complexity that would otherwise be required.

(iii) The complexity of estimation is fit for the task. For persistent navigation, efficient localization of the robot to the precision of a discrete pipe in the network is needed, and facilitated by the reduction of the map to a set of discrete locations. For occasional precise localization of a fault in the pipe, estimation would be more robust (a deficiency of standard methods) and efficient, due to the connected structure of the environment.

(iv) The localization method will be designed for a multi-robot system, which is needed to efficiently monitor a large network.

This work will be developed from theory and evaluated in simulation and using experimental data. Work will be done to compare existing localization methods with those developed, to establish benefits of the method proposed.