Sensor Robustness for Autonomous Mobility
When we talk about robustness in sensing, it is not only about ensuring, certifying, and testing equipment so that it meets technical requirements to a certain degree. In a fast-moving industry like the one we live in today, availability and reliability are taken for granted. However, there is a somewhat higher notion of robustness that includes aspects of a different nature: the robustness of the sensor with respect to its environment. Here, availability is not only about the hardware, but also about the ability to resolve the scene regardless of what that scene may be.
An example of how this robustness can be compromised in a specific case is the following: imagine a traditional camera that may have full availability but not in every environment. This gives rise to a set of situations known as corner cases in which, despite availability and compliance with hardware requirements, the sensor cannot perform its task correctly.
Edge Cases
Corner cases are situations in which one or more characteristics in the sensor inputs prevent the scene from being correctly resolved. Although these situations are, by their nature, infrequent (since sensors that work best in the environment are generally used), they cannot go unnoticed. If we think about the technology required for the autonomous systems of tomorrow, we are not far from achieving it, but the most imminent challenge is to achieve it in such a way that they are always available—in other words, the remaining 1 percent of the 99% is always the hardest.
Robustness by Design: The Sensing Stack
This robustness can be achieved if the sensor stack is chosen correctly. To that end, one should seek as much overlap as possible between the implemented sensors. Ideally, this stack should satisfy the following properties:
- Technological Compatibility: It allows the integration and fusion of the sensors in a native and efficient way.
- Economic Viability: No technological solution can be deployed on a scale if it does not manage to offer a competitive price.
Obviously, there is always the trivial solution in which all sensors are used, but this solution is far from economically viable.
