21 February 2020 in IEEE Robotics and Automation Letters
Localization safety, or integrity risk, is the probability of undetected localization failures and a common aviation performance metric used to verify a minimum accuracy requirement. As autonomous robots become more common, applying integrity risk metrics will be necessary to verify localization performance. This letter introduces a new method, solution separation, to quantify landmark-based mobile robot localization safety for fixed-lag smoothing estimators and compares it's computation time and fault detection capabilities to a chi-squared integrity monitoring method. Results show that solution separation is more computationally efficient and results in a tighter upper-bound on integrity risk when few measurements are included, which makes it the method of choice for lightweight, safety-critical applications such as UAVs. Conversely, chi-squared requires more computing resources but performs better when more measurements are included, making the method more appropriate for high performance computing platforms such as autonomous vehicles.
8 October 2020 in The International Journal of Robotics Research
The problem of quantifying robot localization safety in the presence of undetected sensor faults is critical when preparing for future applications where robots may interact with humans in life-critical situations; however, the topic is only sparsely addressed in the robotics literature. In response, this work leverages prior work in aviation integrity monitoring to tackle the more challenging case of evaluating localization safety in Global Navigation Satellite System (GNSS)-denied environments. Localization integrity risk is the probability that a robot’s pose estimate lies outside pre-defined acceptable limits while no alarm is triggered. In this article, the integrity risk (i.e., localization safety) is rigorously upper bounded by accounting for both nominal sensor noise and other non-nominal sensor faults. An extended Kalman filter is employed to estimate the robot state, and a sequence of innovations is used for fault detection. The novelty of the work includes (1) the use of a time window to limit the number of monitored fault hypotheses while still guaranteeing safety with respect to previously occurring faults and (2) a new method to account for faults in the data association process.