Abstract: This thesis addresses the design, development and implementation of a motion measuring sensor for use in the context of autonomous navigation. The sensor combines both visual and range information in a robust estimation framework. Range is used to allow calculation of translational ground plane velocity, to ensure real-time constraints are met and to provide a simple means of segmenting the environment into coherently moving regions. A prototype sensor has been implemented using Field Programmable Gate Array technology. This has allowed a 'system on a chip' solution with the only external devices being sensors (camera and range) and primary memoly. The sensor can process images of up to 512*32 pixels resolution in realtime. This thesis shows that, in the context of autonomous navigation the concept of real-time is linked to both object dynamics and sensor sampling considerations. Real time is shown to be 16Hz in the test environment used in this thesis. A combination of offline simulation results (using artificially generated data mimicking the real world thus allowing quantitative performance analysis) and real-time experimental results illustrates the performance of our sensor. This thesis makes the following contributions: 1. It presents the design and implementation of an integrated motion sensing solution that utilises both range and vision to robustly estimate rigid, translational ground plane motion for the purpose of autonomous navigation. 2. It develops the concept of dynamic scale space - a technique that utilises assumed environmental dynamics to focus motion estimation on the closest object so that the sensor meets real time requirements. 3. It develops a simple, iterative robust averaging estimator based on the concept of Least Trimmed Squares. This estimator (the Least Trimmed Squared Variant or LISV estimator) does not require reordering of data or stochastic sampling and does not have parameters that must be tuned to suit the data. At every iteration, the LTSV estimator requires a simple update of threshold parameters, a single division plus two addition operations for each data element. The performance of the LTSV estimator is compared against more traditional estimators (least squares, median, least trimmed squared and the Lorentzian M-Estimator) demonstrating its rapid convergence and consistently low bias. The simplicity and rapid convergence of the estimator are achieved at the expense of statistical efficiency. 4. It demonstrates the use of range information as a means of segmenting the environment into regions we call blobs, under the assumption that each blob moves coherently. In the domain of custom hardware implementations of motion estimation, we believe our solution is the first that; 1. uses both range and visual data, 2. estimates motion using a robust estimation frame work and, 3. embeds the motion estimation process in a (dynamic) scale space framework.