The biggest source of uncertainty in autonomous vehicles is simply other traffic participants
Autonomous vehicles engineering is quite challenging due to the complexity of the system being created and the environment where the vehicles have to operate. An autonomous vehicle needs to create a representation of its surroundings, identifying for instance where the vehicle is in the road, traffic signs, static and dynamic entities in the road or on the road vicinity.
The vehicle can use different sensors to collect information about the environment so that it can use this information to take decisions. Perceiving the environment can be achieved by taking millions of distance measurements to any objects in the vicinity of the vehicle, for instance using LIDAR sensors, or capturing images around the vehicle, using image sensors (a.k.a. cameras). The captured information is processed in different ways and used by the vehicle to decide how much it should accelerate, brake, etc.
The most important aspect is that any mistakes in such measurements done in the perception may affect the behavior of the AV. Measuring anything always has certain uncertainty associated to it. The uncertainty can have several different causes, such as failing sensors, improper operating conditions, interference from other sensors, etc.
The problem of uncertainty is that it can directly affect many other components of the autonomous vehicle. For instance, a simple estimation of how far another car is from the AV can be greatly impacted by uncertainties in the perception. The bigger the error in estimating the position of other vehicles on the road, the bigger the chances of collisions.
There is vast literature studying how to how to improve the perception of an autonomous vehicle removing any associated uncertainties. We believe that measuring the distances between the autonomous vehicle and other entities surrounding it is a matter of improving its accuracy and precision.
However, when one is driving one has to take into account not only the static, dynamic objects, vehicles, etc. surrounding the autonomous vehicle but also the intentions of other road users. Basically, when driving we have to assume other drivers will behave in a certain predictable way. The perception system also needs to take into account the behavior of other road users. Due to the different nature of measuring where other road users, obstacles, etc are in the road versus predicting the behavior of other road users. It is possible to clearly distinguish two main types of uncertainty affecting the AV systems as in the Figure below.
The error in measurements of distances to other road users is known BEFORE the fact. The error in predicting the motion of other traffic participants is only know AFTER the fact.
We believe that the uncertainty originating from the Perception is measurable and its accuracy and precision is a matter of engineering. However, the uncertainty originating from the environment, that is the motion of other traffic participants, can not be measured, or predicted with a 100% accuracy, by definition.
Currently, there is no way to “get inside the brain” of another road user (e.g. a driver or a pedestrian) and know exactly what this person wants to do. Because of this the assessment of an autonomous vehicle prediction of the motion of other road users can only be made after the fact.
Because of this the assessment of an autonomous vehicle prediction of the motion of other road users can only be made after the fact.
Concluding
Other traffic participants, such as drivers, pedestrians, cyclists, etc. provide the biggest source of uncertainty ever to an autonomous vehicle. Autonomous vehicles only know about any prediction errors after the fact.
The biggest challenge then is to make safe decisions knowing that the autonomous vehicle have imperfect predictions.
Acknowledgments
Thank you Hoang Tung Dinh, Quentin de Clercq, and Alfredo D’Alava Jr. for contributing with discussions and feedback to this article.
