Abstract: Self-driving cars, presently under on road testing stage, contain software to drive the vehicle without intervention of a human driver. The two major competencies of this software are perception and planning. When any discrepancy is detected in any of these competencies, it causes “failure in operation” or “disengagement” i.e., control is handed over to human driver present as a back-up arrangement. In this paper, on-road testing data is considered for self-driving cars in California (USA) for eight manufacturers whose vehicles have been tested for at least 10,000 miles. The various reasons for reported disengagement or failure have been attributed to error in two software competencies of self-driving cars - perception competency or planning competency. The “number of miles driven” has been taken as a dependent parameter while cumulative failure due to perception or due to planning have been taken as an independent variable. The seven NHPP software reliability failure-count models have been compared to identify the best-fit model for number of miles driven against failure due to perception discrepancy and number of miles driven against failures due to planning discrepancy for the above-mentioned eight self-driving car manufacturers. The type of debugging has been identified based on the nest fit model and future predicted values of the respective failures are estimated.

Key words: self-driving car, autonomous vehicle, non homogeneous poisson process (nhpp), perception discrepancy, planning, discrepancy, software reliability