Guaranteed Bounds for Vehicle Motion State Estimates for automated driving

Research Project, 2021 – 2024

With demands for increased efficiency and reduced energy consumption, we see a development toward heavier and longer vehicle combinations with a total length of just over 30 meters. This means that several trucks are connected, and you get more (up to three) articulation points to keep track of. A good truck driver develops over time a good feeling for how the different parts of the vehicle combination move in different situations, for example how much trailers cut into a curve or how best to drive through a roundabout. With an increasing degree of automation of long combinations, the requirements also increase that the vehicles themselves (at any moment) have good control of the various parts of the vehicle combination regarding, for example, relative positioning, speed and acceleration, i.e., the state of the vehicles.

This is possible by putting different sensors on the vehicles (cameras, LiDAR, GPS, angle sensors, speed sensors, accelerometers, etc.) that provide noisy measurements of the vehicles' movement. These measurements are refined by combining them with knowledge (models) of how the vehicles move to make an improved estimate of the state of the various vehicles, a so-called state estimate. The problem is that, all sensors have errors (which depend on the sensor's precision level, noise, etc.) and that models of the vehicle's motion are simplified, which in itself gives rise to misinterpretation of the measured values ​​from the sensors. In addition, these sources of error are affected by external circumstances when, for example, cameras work differently in light and dark, and vehicles move very differently when it is slippery. These senor errors and modelling imperfections give rise to loss of estimation precision, which is difficult for the current system to describe reliably. We can, thus, end up in situations where we have significant errors regarding, e.g., the position of the truck's various parts (estimation error) and, what is even worse, be incorrectly certain that these erroneous estimates are accurate (overconfident).

For automation of long vehicle combinations, these types of errors and uncertainties, even though they may seem relatively small, can have major consequences. For example, an error of one degree between the vehicles in the front part of the vehicle combination can give rise to an incorrect estimate of the lateral position of the rear vehicles of around half a meter. There is thus a great risk of serious incidents with, for example, oncoming traffic or when cornering. Thus, a prerequisite for safe autonomous driving with long vehicle combinations is to develop estimation methods that can reliably capture the uncertainty in its estimates for these complex systems. Only then will the system get a good "feeling" for where each vehicle in the combination is, and the system can, like a human driver, adapt how to control the vehicles accordingly.

In this project, we will study how to improve the uncertainty description of the state of vehicles so that we can realize safe automated driving of long vehicle combinations in complex traffic environments. We plan to take a holistic approach to the problem by approaching the problem from three different angles: 1) identify effective measures on adjacent subsystems that simplify the problem, such as requirements on sensors and actuators such that we can describe these, to a greater extent and with greater validity, with simpler models. As a consequence, we will reduce the errors and uncertainties in the state estimates, but will at the same time lead to higher requirements for sensors and actuators, which drives costs and the need for verification. 2) develop new/existing estimation methods where we not only focus on estimation precision (state) but also place great emphasis on its uncertainty description. 3) explore opportunities to use data (machine learning) to train methods/models that can describe the complex nature of the system and its uncertainties in a better way than we can with (hand-designed) methods and models. How to train these data-driven models to produce good uncertainty descriptions is generally an important research question as the description of uncertainty from these is often unsatisfactory.