Identifying Admissible Uncertainty Bounds for the Input of Planning Algorithms

In automated driving, a strong relation between the sensing modules and the planning module exists, since the former provide uncertain information to the latter. To ensure good data processing over the whole sense-plan-act chain, the quality of the planner’s input needs to be analyzed: How much uncertainty can the planning module handle while still providing reliable decisions, and are the sensing modules capable of ensuring this quality for the information they provide? We present a method that identifies uncertainty bounds for every input such that with a given probability changes in the input, e.g., due to measurement errors, do not lead to unacceptably large deviations in the planning result. For this, we define a stochastic optimization program, which characterizes non-influential input parameters with a stochastic constraint, and give a discrete approximation to the problem. To solve it, we extend a multilevel coordinate search algorithm by a sensitivity analysis. Finally, we determine accuracy requirements for a longitudinal-lateral driver model in a roundabout scenario. We see that the proposed method is capable of determining situation- and planner-dependent bounds. With this, we are able to detect weaknesses of the planning algorithm offline as well as to examine if a more conservative driving maneuver is required due to a lack of information.