Stochastic techniques for the cooperative navigation of autonomous mobile robots

The cooperative navigation system (CNS) algorithm is based on a Kalman filter which uses inter-robot relative-displacement measurements to improve the collective position estimates of a group of autonomous mobile robots. The principal motivation is reduction of the unbounded growth of position errors when dead-reckoning is used as the sole means of navigation for individual robots. In this thesis, dynamics-based models of groups of two-drivewheel robots are created and used in extensive simulations to verify the navigation improvement properties of the CNS algorithm. Observability properties of such groups of robots are also proven, and used to predict the scenarios in which the navigation errors remain bounded, as well as those wherein the errors are unbounded. A simple quantitative relationship has been found for the lower bound of error reduction with a CNS group composed of mobile robots moving in formation. Simulations (of multiple robots) and experiments (with a pair of small two-drivewheel robots) reveal that this lower bound can be approached under the right conditions even with delays between inter-robot measurements. A description of two hardware development projects essential for the experimental part of the research is also incorporated in the thesis. The first was the creation of an instrument platform for the rangefinder robot that allows it to obtain the correct range and bearing of the beacon-carrying reference robot, while ignoring other objects in the environment. The second hardware development task was the creation of a controller for the rangefinder robot consisting of two Texas Instruments (TI) digital signal processors (DSPs): a TI TMS320F243 and TMS320F2812 in a shared-memory configuration. The dual DSP controller permits the rangefinder robot to continuously perform several simultaneous tasks in real time, including the control of its motion and of its sensor platform, and the carrying out of Kalman filter calculations to estimate 18 state variables every 50 ms. Consequently, replication and adaptation of the hardware is now straightforward for further work and development in this area.