Sensor Data Integrity and Mitigation of Perceptual Failures

This document is a progress report for the second year of the research project “Sensor Data Integrity and Mitigation of Perceptual Failures”, sponsored by the AFOSR/AOARD under agreement number FA2386-10-1-4153. The objective of this research is to better understand and promote integrity and dependability of unmanned ground vehicles (UGVs), with a focus on their perceptual systems. This will provide UGVs with the ability to achieve long-term autonomous operations in o↵-road environments, including in challenging conditions. We investigate methods to mitigate, detect and/or recover from perceptual failures and failures due to perception. The document is divided into four main sections. First, we further develop the concept of laser-to-radar sensing redundancy for resilient perception in adverse environmental conditions (e.g. presence of smoke, airborne dust or heavy rain). We show that we can separate laser points due to dust or smoke clouds from the points corresponding to actual dense obstacles using a consistency test with data acquired by a mm-wave radar. A direct benefit is that the UGV does not misinterpret dust/smoke clouds or heavy rain particles as actual obstacles. A more general benefit is that using this technique, a UGV can keep building accurate environment models in clear environmental conditions, while maintaining resilient perception in adverse environmental conditions. Second, we analyse the influence of di↵erent terrain geometry representations (in particular state-of-the-art techniques based on Gaussian Process regression) on the estimation of traversability of a ground vehicle. In particular, we discuss how these terrain geometry modelling techniques may whether mitigate or generate errors in the estimation of traversability. This preliminary study will drive further investigation into traversability estimation techniques that are more accurate and more robust to occlusions and sensing errors. In the third part, we further develop the concept of resilient navigation through probabilistic modality reconfiguration that we introduced in the previous progress report in 2011. The update includes new experiments, and a comparison with a simpler thresholding technique to decide the most appropriate navigation modality. The benefit of this navigation modality reconfiguration technique is the online mitigation of, or recovery from, unpredictable errors such as control deviations, map failures and localisation faults. Finally, in the last part of this document pushes the study to a higher level in the architecture