This research will develop new navigation integrity evaluation methods to answer the fundamental research question, how can we prove that a co-robotic, self-driving car is safer than a human driver? Integrity is a measure of trust in a sensor’s information, and has been used for decades to guarantee safety in life-critical aviation systems. Extending this idea to self-driving cars holds significant challenges because, in contrast with aviation navigation, self-driving cars require sensor data other than GPS and operate in a relatively unpredictable, constantly changing environment. The results of this work will create a high-level, sensor-independent, quantifiable metric that can be used to compare, evaluate, and certify safety across self-driving car manufacturers. We expect the results will be part of the larger research effort to enable self-driving cars to reduce traffic, reduce emissions, generate enormous potential cost savings, and, most importantly, reduce fatalities. Furthermore, we plan to disseminate this work through open access to our experimental data as well as using our ongoing relationship with Chicago’s Museum of Science and Industry to demonstrate our results to the public. The latter will include a hands-on demonstration during National Robotics Week that illustrates how navigation safety is impacted by sensors, vehicle dynamics, and the environment.
Specifically, this work will provide new experimental and analytical methods to quantify and prove self-driving car safety. While developing these methods, we will advance knowledge in several previously-unexplored areas by being the first to: 1) establish high-integrity sensor measurement error and fault models for non-GPS sensors, 2) develop analytical methods to quantify the safety risk of feature extraction and data association algorithms required in lidar, radar, and camera-based localization, 3) design multi-sensor pose estimators and integrity monitors to evaluate the impact of undetected sensor faults on safety risk, and 4) derive, analyze, and experimentally implement integrity risk prediction methods in dynamic environments.
We are collaborating with Prof. Mathieu Joerger at the University of Arizona on this project.
Specifically, this work will provide new experimental and analytical methods to quantify and prove self-driving car safety. While developing these methods, we will advance knowledge in several previously-unexplored areas by being the first to: 1) establish high-integrity sensor measurement error and fault models for non-GPS sensors, 2) develop analytical methods to quantify the safety risk of feature extraction and data association algorithms required in lidar, radar, and camera-based localization, 3) design multi-sensor pose estimators and integrity monitors to evaluate the impact of undetected sensor faults on safety risk, and 4) derive, analyze, and experimentally implement integrity risk prediction methods in dynamic environments.
We are collaborating with Prof. Mathieu Joerger at the University of Arizona on this project.